Prior work has shown that vector-based pseudo relevance feedback (PRF) is an effective technique for query expansion for improving retrieval results in dense information retrieval. In dense retrieval, ColBERT-PRF has emerged as a novel mechanism, using cluster centroids built from feedback documents as PRF expansion tokens and leveraging statistical information from the closest neighboring token ids to dictate how useful these expansion tokens are. While this approach has been shown to work well in the monolingual retrieval setting for English using the original ColBERT infrastructure, such systems have since evolved to improve inference speed, reduce storage and memory usage, and support cross-language (CLIR) and multilingual (MLIR) retrieval. As a result, many of these advancements have reduced the ability to utilize token-level statistics. In this work, we aim to explore how well this type of approach can adapt to dense retrieval models when it is not feasible to use surface-form information to pick discriminating expansion tokens. Furthermore, we explore alternative clustering mechanisms, such as HDBScan, to compare how different clustering methods perform at building clusters that can be useful for PRF. Experiments on MLIR, CLIR, and Report Generation tasks, such as those in the TREC 2024 NeuCLIR Report Generation Pilot Task, show that even without access to these token statistics, the use of cluster centroids for PRF can still improve nDCG and α-nDCG by up to 12%.

Most current zero-shot voice conversion methods rely on externally supervised components, particularly speaker encoders, for training. To explore alternatives that eliminate this dependency, this paper introduces GenVC, a novel framework that disentangles speaker identity and linguistic content from speech signals in a self-supervised manner. GenVC leverages speech tokenizers and an autoregressive, Transformer-based language model as its backbone for speech generation. This design supports large-scale training while enhancing both source speaker privacy protection and target speaker cloning fidelity. Experimental results demonstrate that GenVC achieves notably higher speaker similarity, with naturalness on par with leading zero-shot approaches. Moreover, due to its autoregressive formulation, GenVC introduces flexibility in temporal alignment, reducing the preservation of source prosody and speaker-specific traits, and making it highly effective for voice anonymization.11Audio samples, code, and model checkpoints are available at https://caizexin.github.io/GenVC/index.html

The Recurrent Neural Network-Transducer (RNN-T) is widely adopted in end-to-end (E2E) automatic speech recognition (ASR) tasks but depends heavily on large-scale, high-quality annotated data, which are often costly and difficult to obtain. To mitigate this reliance, we propose a Weakly Supervised Transducer (WST), which integrates a flexible training graph designed to robustly handle errors in the transcripts without requiring additional confidence estimation or auxiliary pre-trained models. Empirical evaluations on synthetic and industrial datasets reveal that WST effectively maintains performance even with transcription error rates of up to 70%, consistently outperforming existing Connectionist Temporal Classification (CTC)-based weakly supervised approaches, such as Bypass Temporal Classification (BTC) and Omni-Temporal Classification (OTC). These results demonstrate the practical utility and robustness of WST in realistic ASR settings. The implementation will be publicly available.

We address the challenge of detecting synthesized speech under distribution shifts—arising from unseen synthesis methods, speakers, languages, or audio conditions—relative to the training data. Fewshot learning methods are a promising way to tackle distribution shifts by rapidly adapting on the basis of a few in-distribution samples. We propose a self-attentive prototypical network to enable more robust fewshot adaptation. To evaluate our approach, we systematically compare the performance of traditional zero-shot detectors and the proposed fewshot detectors, carefully controlling training conditions to introduce distribution shifts at evaluation time. In conditions where distribution shifts hamper the zero-shot performance, our proposed few-shot adaptation technique can quickly adapt using as few as $\mathbf{1 0}$ in-distribution samples—achieving upto 32% relative EER reduction on deepfakes in Japanese language and 20% relative reduction on ASVspoof 2021 Deepfake dataset.

We present MADASR 2.0, a challenge at ASRU 2025 aimed at advancing multilingual and multidialectal automatic speech recognition (ASR) in low-resource Indian languages. Building on the 2023 edition, it introduces a subset of the RESPIN corpus, over 1200 hours of read speech across 8 languages and 33 dialects, with test sets including both read and spontaneous speech. The challenge comprises four tracks varying by training data size and external resource usage, and supports auxiliary tasks like language and dialect identification. We detail the dataset, tasks, baselines, and submissions and analyse trends across tracks and speech styles. Results highlight the continued difficulty of spontaneous ASR, the benefits of multitask and transfer learning, and effective strategies for building dialect-aware ASR systems. MADASR 2.0 offers a standardised benchmark to support future research on inclusive and scalable ASR for linguistically diverse populations.

The success of large language models has driven interest in developing similar speech processing capabilities. However, a key challenge is the scarcity of high-quality spontaneous speech data, as most existing datasets contain scripted dialogues. To address this, we present a novel pipeline for eliciting and recording natural dialogues and release our dataset with 100+ hours of spontaneous speech. Our approach fosters fluid, natural conversations while encouraging a diverse range of topics and interactive exchanges. Unlike traditional methods, it facilitates genuine interactions, providing a reproducible framework for future data collection. This paper introduces our dataset and methodology, laying the groundwork for addressing the shortage of spontaneous speech data. We plan to expand this dataset in future stages, offering a growing resource for the research community.

We propose a method to scale accented TTS training to large, accent-diverse datasets that often lack consistent, high-quality accent labels. Our approach relies on a speech geolocation model to infer accent labels directly from audio. To improve speaker generalization and encourage disentangling speaker from accent we explore timbre augmentation through kNN voice conversion. We validate our approach on CommonVoice by fine-tuning XTTS-v2 with accent labels inferred or improved via geolocation. According to various automated metrics based on embeddings extracted from an accent identification model, the resulting accented TTS model produces speech with better accent fidelity compared to XTTS-v2 fine-tuned on self-reported accent labels in CommonVoice, or other existing accented TTS models. According to human evaluation, it was clear that the geolocation model based data discovery and enhancement improved the naturalness and accent fidelity of generated speech. However, the effect of different data augmentation strategies was less clear.

Cihan Xiao, Matthew Wiesner, Debashish Chakraborty, Reno Kriz, Keith Cunningham, Kenton Murray, Kevin Duh, Luis Tavarez-Arce, Paul McNamee, Sanjeev Khudanpur. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

We propose a novel approach to enable the use of large, single-speaker ASR models, such as Whisper, for target speaker ASR. The key claim of this method is that it is much easier to model relative differences among speakers by learning to condition on frame-level diarization outputs than to learn the space of all speaker embeddings. We find that adding even a single bias term per diarization output type before the first transformer block can transform single-speaker ASR models into target-speaker ASR models. Our approach also supports speaker-attributed ASR by sequentially generating transcripts for each speaker in a diarization output. This simplified method outperforms baseline speech separation and diarization cascade by 12.9 % absolute ORC-WER on the NOTSOFAR-1 dataset.

We describe our submission to the 2024 VoicePrivacy Attacker Challenge. We propose three main categories of methods to improve ASV performance against anonymized speech: improvements to the underlying classifier, alternative distance metrics when computing ASV scores, and kNN-VC normalization. By simultaneously employing one or more of these methods, we were able to achieve a significant reduction in EER against all of the submitted anonymization systems in the VoicePrivacy Challenge.

Yining Lu, Dixuan Wang, Tianjian Li, Dongwei Jiang, Sanjeev Khudanpur, Meng Jiang, Daniel Khashabi. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Knowing the particular context associated with a conversation can help improving the performance of an automatic speech recognition (ASR) system. For example, if we are provided with a list of in-context words or phrases — such as the speaker’s contacts or recent song playlists — during inference, we can bias the recognition process towards this list. There are many works addressing contextual ASR; however, there is few publicly available real benchmark for evaluation, making it difficult to compare different solutions. To this end, we provide a corpus (“ConEC”) and baselines to evaluate contextual ASR approaches, grounded on real-world applications. The ConEC corpus is based on public-domain earnings calls (ECs) and associated supplementary materials, such as presentation slides, earnings news release as well as a list of meeting participants’ names and affiliations. We demonstrate that such real contexts are noisier than artificially synthesized contexts that contain the ground truth, yet they still make great room for future improvement of contextual ASR technology

Connectionist temporal classification (CTC) models are known to have peaky output distributions. Such behavior is not a problem for automatic speech recognition (ASR), but it can cause inaccurate forced alignments (FA), especially at finer granularity, e.g., phoneme level. This paper aims at alleviating the peaky behavior for CTC and improve its suitability for forced alignment generation, by leveraging label priors, so that the scores of alignment paths containing fewer blanks are boosted and maximized during training. As a result, our CTC model produces less peaky posteriors and is able to more accurately predict the offset of the tokens besides their onset. It outperforms the standard CTC model and a heuristics-based approach for obtaining CTC’s token offset timestamps by 12 − 40% in phoneme and word boundary errors (PBE and WBE) measured on the Buckeye and TIMIT data. Compared with the most widely used FA toolkit Montreal Forced Aligner (MFA), our method performs similarly on PBE/WBE on Buckeye, yet falls behind MFA on TIMIT. Nevertheless, our method has a much simpler training pipeline and better runtime efficiency. Our training recipe and pretrained model are released in TorchAudio.

Incorporating longer context has been shown to benefit machine translation, but the inclusion of context in end-to-end speech translation (E2E-ST) remains under-studied. To bridge this gap, we introduce target language context in E2E-ST, enhancing coherence and overcoming memory constraints of extended audio segments. Additionally, we propose context dropout to ensure robustness to the absence of context, and further improve performance by adding speaker information. Our proposed contextual E2E-ST outperforms the isolated utterance-based E2E-ST approach. Lastly, we demonstrate that in conversational speech, contextual information primarily contributes to capturing context style, as well as resolving anaphora and named entities.

Designing effective automatic speech recognition (ASR) systems for Code-Switching (CS) often depends on the availability of the transcribed CS resources. To address data scarcity, this paper introduces Speech Collage, a method that synthesizes CS data from monolingual corpora by splicing audio segments. We further improve the smoothness quality of audio generation using an overlap-add approach. We investigate the impact of generated data on speech recognition in two scenarios: using in-domain CS text and a zero-shot approach with synthesized CS text. Empirical results highlight up to 34.4% and 16.2% relative reductions in Mixed-Error Rate and Word-Error Rate for in-domain and zero-shot scenarios, respectively. Lastly, we demonstrate that CS augmentation bolsters the model’s code-switching inclination and reduces its monolingual bias.

Languages usually switch within a multilingual speech signal, especially in a bilingual society. This phenomenon is referred to as code-switching (CS), making automatic speech recognition (ASR) challenging under a multilingual scenario. We propose to improve CS-ASR by biasing the hybrid CTC/attention ASR model with multi-level language information comprising frame-and token-level language posteriors. The interaction between various resolutions of language biases is subsequently explored in this work. We conducted experiments on datasets from the ASRU 2019 code-switching challenge. Compared to the baseline, the proposed interactive language biases (ILB) method achieves higher performance and ablation studies highlight the effects of different language biases and their interactions. In addition, the results presented indicate that language bias implicitly enhances internal language modeling, leading to performance degradation after employing an external language model.

Nathaniel Robinson, Raj Dabre, Ammon Shurtz, Rasul Dent, Onenamiyi Onesi, Claire Monroc, Loïc Grobol, Hasan Muhammad, Ashi Garg, Naome Etori, Vijay Murari Tiyyala, Olanrewaju Samuel, Matthew Stutzman, Bismarck Odoom, Sanjeev Khudanpur, Stephen Richardson, Kenton Murray. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Advances in speech technology now allow unprecedented access to personally identifiable information through speech. To protect such information, the differential privacy field has explored ways to anonymize speech while preserving its utility, including linguistic and paralinguistic aspects. However, anonymizing speech while maintaining emotional state remains challenging. We explore this problem in the context of the VoicePrivacy 2024 challenge. Specifically, we developed various speaker anonymization pipelines and find that approaches either excel at anonymization or preserving emotion state, but not both simultaneously. Achieving both would require an in-domain emotion recognizer. Additionally, we found that it is feasible to train a semi-effective speaker verification system using only emotion representations, demonstrating the challenge of separating these two modalities.

Spatial information is a critical clue for multi-channel multispeaker target speech recognition. Most state-of-the-art multi-channel Automatic Speech Recognition (ASR) systems extract spatial features only during the speech separation stage, followed by standard single-channel ASR on the separated speech. This approach results in an inefficient, lengthy pipeline and sub-optimal ASR performance due to the accumulated errors from preprocessing modules. Furthermore, most spatial feature extraction methods depend on the knowledge of speaker positions and microphone topology, making the systems reliant on specific settings and challenging to adapt to new equipment. In this work, we propose a solution to these issues with a lightweight embedding module named SpatialEmb, which extracts and encodes spatial information directly for the ASR model, supporting both fixed and arbitrary microphone topology. We conduct comprehensive experiments on AliMeeting, a real meeting corpus, to determine the optimal model design for SpatialEmb in terms of both performance and efficiency. Our best model trained with 105 hours Train-Ali-far achieves 17.04% and 20.32% character error rates (CER) on the Eval and Test sets, establishing a new state-of-the-art result with the same training data.

Poisoning backdoor attacks involve an adversary manipulating the training data to induce certain behaviors in the victim model by inserting a trigger in the signal at inference time. We adapted clean label backdoor (CLBD)-data poisoning attacks, which do not modify the training labels, on state-of-the-art speech recognition models that support/perform a Spoken Language Understanding task, achieving 99.8% attack success rate by poisoning 10% of the training data. We analyzed how varying the signal-strength of the poison, percent of samples poisoned, and choice of trigger impact the attack. We also found that CLBD attacks are most successful when applied to training samples that are inherently hard for a proxy model. Using this strategy, we achieved an attack success rate of 99.3% by poisoning a meager 1.5% of the training data. Finally, we applied two previously developed defenses against gradient-based attacks, and found that they attain mixed success against poisoning.

Training automatic speech recognition (ASR) systems requires large amounts of well-curated paired data. However, human annotators usually perform “non-verbatim” transcription, which can result in poorly trained models. In this paper, we propose Omni-temporal Classification (OTC), a novel training criterion that explicitly incorporates label uncertainties originating from such weak supervision. This allows the model to effectively learn speech-text alignments while accommodating errors present in the training transcripts. OTC extends the conventional CTC objective for imperfect transcripts by leveraging weighted finite state transducers. Through experiments conducted on the LibriSpeech and LibriVox datasets, we demonstrate that training ASR models with OTC avoids performance degradation even with transcripts containing up to 70% errors, a scenario where CTC models fail completely. Our implementation is available at https://github.com/k2-fsa/icefall.

Our novel technique utilizes a Joint Energy-based Model (JEM) that integrates both discriminative and generative approaches to increase resistance against dirty-label backdoor attacks. Our approach is especially effective when the trigger is short or hardly perceivable. We simulate the attack on the Speech Commands Dataset consisting of 1s audio clips. During training, we use JEM to model a view of the input implemented by a randomly selected 610ms window. During inference, we combine all (40) possible views utilizing a generative part of JEM. The resulting system has slightly decreased accuracy but significantly increased resistance shown in multiple scenarios. Interestingly, replacing JEM with a standard discriminative model (Disc) provides increased resistance with a lesser effect compared to JEM but maintains accuracy. We introduce an extension motivated by semi-supervised training that further improves JEM but not Disc. JEM can also benefit from Gaussian noise during evaluation.

Poisoning attacks entail attackers intentionally tampering with training data. In this paper, we consider a dirty-label poisoning attack scenario on a speech commands classification system. The threat model assumes that certain utterances from one of the classes (source class) are poisoned by superimposing a trigger on it, and its label is changed to another class selected by the attacker (target class). We propose a filtering defense against such an attack. First, we use DIstillation with NO labels (DINO) to learn unsupervised representations for all the training examples. Next, we use K-means and LDA to cluster these representations. Finally, we keep the utterances with the most repeated label in their cluster for training and discard the rest. For a 10% poisoned source class, we demonstrate a drop in attack success rate from 99.75% to 0.25%. We test our defense against a variety of threat models, including different target and source classes, as well as trigger variations.

This paper describes the ESPnet Unsupervised ASR Open-source Toolkit (EURO), an end-to-end open-source toolkit for unsupervised automatic speech recognition (UASR). EURO adopts the state-of-the-art UASR learning method introduced by the Wav2vec-U, originally implemented at FAIRSEQ, which leverages self-supervised speech representations and adversarial training. In addition to wav2vec2, EURO extends the functionality and promotes reproducibility for UASR tasks by integrating S3PRL and k2, resulting in flexible frontends from 27 self-supervised models and various graph-based decoding strategies. EURO is implemented in ESPnet and follows its unified pipeline to provide UASR recipes with a complete setup. This improves the pipeline’s efficiency and allows EURO to be easily applied to existing datasets in ESPnet. Extensive experiments on three mainstream self-supervised models demonstrate the toolkit’s effectiveness and achieve state-of-the-art UASR performance on TIMIT and LibriSpeech datasets. EURO will be publicly available at https://github.com/espnet/espnet, aiming to promote this exciting and emerging research area based on UASR through open-source activity.

Self-supervised learning (SSL) methods which learn representations of data without explicit supervision have gained popularity in speech-processing tasks, particularly for single-talker applications. However, these models often have degraded performance for multi-talker scenarios — possibly due to the domain mismatch — which severely limits their use for such applications. In this paper, we investigate the adaptation of upstream SSL models to the multi-talker automatic speech recognition (ASR) task under two conditions. First, when segmented utterances are given, we show that adding a target speaker extraction (TSE) module based on enrollment embeddings is complementary to mixture-aware pre-training. Second, for unsegmented mixtures, we propose a novel joint speaker modeling (JSM) approach, which aggregates information from all speakers in the mixture through their embeddings. With controlled experiments on Libri2Mix, we show that using speaker embeddings provides relative WER improvements of 9.1% and 42.1% over strong baselines for the segmented and unsegmented cases, respectively. We also demonstrate the effectiveness of our models for real conversational mixtures through experiments on the AMI dataset. Our code and models are open-sourced on https://github.com/HuangZiliAndy/SSL_for_multitalker.

Keyword search (KWS) systems are commonly built on top of existing automatic speech recognition (ASR) systems. However, end-to-end (E2E) ASR models are not naturally equipped with word-level timing information or confidence. Existing methods for re-purposing E2E ASR systems for KWS are largely heuristic or model-specific. In this paper, we describe a general KWS pipeline, applicable to any ASR model that generates N-best lists. We extract timing information using either external word-aligners, or time-preserving weighted finite-state transducer-based decoders. We show that our light-weight, ASR-agnostic approach for confidence estimation based on N-best lists outperforms other commonly used heuristics, such as using the decoder’s softmax probability, and even a more complicated dedicated confidence estimation model (CEM). Finally, we compare our performance to hybrid ASR models, extensively evaluating the impact of word-level timing, confidence, and recall on KWS performance. Our KWS pipeline is available online1, suitable for evaluating the aforementioned ASR components as downstream tasks.

Code-switching (CS) occurs when languages switch within a speech signal and leads to language confusion for automatic speech recognition (ASR). We address the problem of language confusion for improving CS-ASR from two perspectives: incorporating and disentangling language information. We incorporate language information within the CS-ASR model by dynamically biasing the model with token-level language posteriors corresponding to outputs of a sequence-to-sequence auxiliary language diarization (LD) module. In contrast, the disentangling process reduces the difference between languages via adversarial training so as to normalize two languages. We conduct experiments on the SEAME dataset. Compared to the baseline model, both the joint optimization with LD and the language posterior bias achieve performance improvement. Comparison of the proposed methods indicates that incorporating language information is more effective than disentangling for reducing language confusion in CS speech.

To enhance the reliability and robustness of language identification (LID) and language diarization (LD) systems for heterogeneous populations and scenarios, there is a need for speech processing models to be trained on datasets that feature diverse language registers and speech patterns. We present the MERLIon CCS challenge, featuring a first-of-its-kind Zoom video call dataset of parent-child shared book reading, of over 30 hours with over 300 recordings, annotated by multilingual transcribers using a high-fidelity linguistic transcription protocol. The audio corpus features spontaneous and in-the-wild English-Mandarin code-switching, child-directed speech in non-standard accents with diverse language-mixing patterns recorded in a variety of home environments. This report describes the corpus, as well as LID and LD results for our baseline and several systems submitted to the MERLIon CCS challenge using the corpus.

Language development experts need tools that can automatically identify languages from fluent, conversational speech and provide reliable estimates of usage rates at the level of an individual recording. However, LID systems are typically evaluated on metrics such as equal error rate and balanced accuracy, applied at the level of an entire speech corpus. These overview metrics do not provide information about model performance at the level of individual speakers, recordings, or units of speech with different linguistic characteristics. Overview statistics may mask systematic errors in model performance for some subsets of the data, and consequently, have worse performance on data derived from some subsets of human speakers, creating a kind of algorithmic bias. Here, we investigate how well a number of LID systems perform on individual recordings and speech units with different linguistic properties in the MERLIon CCS Challenge featuring accented code-switched child-directed speech.

The Streaming Unmixing and Recognition Transducer (SURT) model was proposed recently as an end-to-end approach for continuous, streaming, multi-talker speech recognition (ASR). Despite impressive results on multi-turn meetings, SURT has notable limitations: (i) it suffers fromleakageandomissionrelated errors; (ii) it is computationally expensive, due to which it has not seen adoption in academia; and (iii) it has only been evaluated on synthetic mixtures. In this work, we propose several modifications to the original SURT which are carefully designed to fix the above limitations. In particular, we (i) change the unmixing module to a mask estimator that uses dual-path modeling, (ii) use a streaming zipformer encoder and a stateless decoder for the transducer, (iii) perform mixture simulation using force-aligned subsegments, (iv) pre-train the transducer on single-speaker data, (v) use auxiliary objectives in the form of masking loss and encoder CTC loss, and (vi) perform domain adaptation for far-field recognition. We show that our modifications allow SURT 2.0 to outperform its predecessor in terms of multi-talker ASR results, while being efficient enough to train with academic resources. We conduct our evaluations on 3 publicly available meeting benchmarks — LibriCSS, AMI, and ICSI, where our best model achieves WERs of 16.9%, 44.6% and 32.2%, respectively, on far-field unsegmented recordings. We release training recipes and pre-trained models:https://sites.google.com/view/surt2.

Speech enhancement and separation are two fundamental tasks for robust speech processing. Speech enhancement suppresses background noise while speech separation extracts target speech from interfering speakers. Despite a great number of supervised learning-based enhancement and separation methods having been proposed and achieving good performance, studies on applying self-supervised learning (SSL) to enhancement and separation are limited. In this paper, we evaluate 13 SSL upstream methods on speech enhancement and separation downstream tasks. Our experimental results on Voicebank-DEMAND and Libri2Mix show that some SSL representations consistently outperform baseline features including the short-time Fourier transform (STFT) magnitude and log Mel filterbank (FBANK). Furthermore, we analyze the factors that make existing SSL frameworks difficult to apply to speech enhancement and separation and discuss the representation properties desired for both tasks. Our study is included as the official speech enhancement and separation downstreams for SUPERB.

Self-supervised model pretraining has recently garnered significant interest. However, using additional resources in fine-tuning these models has received less attention. We demonstrate how universal phoneset acoustic models can leverage cross-lingual supervision to improve transfer of pretrained self-supervised representations to new languages. We also show how target-language text can be used to enable and improve fine-tuning with the lattice-free maximum mutual information (LF-MMI) objective. In three low-resource languages these techniques greatly improved few-shot learning performance.

We propose a novel model to hierarchically incorporate phoneme and phonotactic information for language identification (LID) without requiring phoneme annotations for training.In this model, named PHO-LID, a self-supervised phoneme segmentation task and a LID task share a convolutional neural network (CNN) module, which encodes both language identity and sequential phonemic information in the input speech to generate an intermediate sequence of "phonotactic" embeddings.These embeddings are then fed into transformer encoder layers for utterance-level LID.We call this architecture CNN-Trans.We evaluate it on AP17-OLR data and the MLS14 set of NIST LRE 2017, and show that the PHO-LID model with multitask optimization exhibits the highest LID performance among all models, achieving over 40% relative improvement in terms of average cost on AP17-OLR data compared to a CNN-Trans model optimized only for LID.The visualized confusion matrices imply that our proposed method achieves higher performance on languages of the same cluster in NIST LRE 2017 data than the CNN-Trans model.A comparison between predicted phoneme boundaries and corresponding audio spectrograms illustrates the leveraging of phoneme information for LID.

The pervasiveness of intra-utterance code-switching (CS) in spoken content requires that speech recognition (ASR) systems handle mixed language. Designing a CS-ASR system has many challenges, mainly due to data scarcity, grammatical structure complexity, and domain mismatch. The most common method for addressing CS is to train an ASR system with the available transcribed CS speech, along with monolingual data. In this work, we propose a zero-shot learning methodology for CS-ASR by augmenting the monolingual data with artificially generating CS text. We based our approach on random lexical replacements and Equivalence Constraint (EC) while exploiting aligned translation pairs to generate random and grammatically valid CS content. Our empirical results show a 65.5% relative reduction in language model perplexity, and 7.7% in ASR WER on two ecologically valid CS test sets. The human evaluation of the generated text using EC suggests that more than 80% is of adequate quality.

We introduce asynchronous dynamic decoder, which adopts an efficient A* algorithm to incorporate big language models in the one-pass decoding for large vocabulary continuous speech recognition. Unlike standard one-pass decoding with on-the-fly composition decoder which might induce a significant computation overhead, the asynchronous dynamic decoder has a novel design where it has two fronts, with one performing "exploration" and the other "backfill". The computation of the two fronts alternates in the decoding process, resulting in more effective pruning than the standard one-pass decoding with an on-the-fly composition decoder. Experiments show that the proposed decoder works notably faster than the standard one-pass decoding with on-the-fly composition decoder, while the acceleration will be more obvious with the increment of data complexity.

This paper proposes a parallel computation strategy and a posterior-based lattice expansion algorithm for efficient lattice rescoring with neural language models (LMs) for automatic speech recognition. First, lattices from first-pass decoding are expanded by the proposed posterior-based lattice expansion algorithm. Second, each expanded lattice is converted into a minimal list of hypotheses that covers every arc. Each hypothesis is constrained to be the best path for at least one arc it includes. For each lattice, the neural LM scores of the minimal list are computed in parallel and are then integrated back to the lattice in the rescoring stage. Experiments on the Switchboard dataset show that the proposed rescoring strategy obtains comparable recognition performance and generates more compact lattices than a competitive baseline method. Furthermore, the parallel rescoring method offers more flexibility by simplifying the integration of PyTorch-trained neural LMs for lattice rescoring with Kaldi.

As the performance of single-channel speech separation systems has improved, there has been a desire to move to more challenging conditions than the clean, near-field speech that initial systems were developed on. When training deep learning separation models, a need for ground truth leads to training on synthetic mixtures. As such, training in noisy conditions requires either using noise synthetically added to clean speech, preventing the use of in-domain data for a noisy-condition task, or training using mixtures of noisy speech, requiring the network to additionally separate the noise. We demonstrate the relative inseparability of noise and that this noisy speech paradigm leads to significant degradation of system performance. We also propose an SI-SDR–inspired training objective that tries to exploit the inseparability of noise to implicitly partition the signal and discount noise separation errors, enabling the training of better separation systems with noisy oracle sources.

Modern wake word detection systems usually rely on neural networks for acoustic modeling. Transformers has recently shown superior performance over LSTM and convolutional networks in various sequence modeling tasks with their better temporal modeling power. However it is not clear whether this advantage still holds for short-range temporal modeling like wake word detection. Besides, the vanilla Transformer is not directly applicable to the task due to its non-streaming nature and the quadratic time and space complexity. In this paper we explore the performance of several variants of chunk-wise streaming Transformers tailored for wake word detection in a recently proposed LF-MMI system, including looking-ahead to the next chunk, gradient stopping, different positional embedding methods and adding same-layer dependency between chunks. Our experiments on the Mobvoi wake word dataset demonstrate that our proposed Transformer model outperforms the baseline convolution network by 25% on average in false rejection rate at the same false alarm rate with a comparable model size, while still maintaining linear complexity w.r.t. the sequence length.

This paper introduces GigaSpeech, an evolving, multi-domain English speech recognition corpus with 10,000 hours of high quality labeled audio suitable for supervised training, and 40,000 hours of total audio suitable for semi-supervised and unsupervised training.Around 40,000 hours of transcribed audio is first collected from audiobooks, podcasts and YouTube, covering both read and spontaneous speaking styles, and a variety of topics, such as arts, science, sports, etc.A new forced alignment and segmentation pipeline is proposed to create sentence segments suitable for speech recognition training, and to filter out segments with low-quality transcription.For system training, GigaSpeech provides five subsets of different sizes, 10h, 250h, 1000h, 2500h, and 10000h.For our 10,000-hour XL training subset, we cap the word error rate at 4% during the filtering/validation stage, and for all our other smaller training subsets, we cap it at 0%.The DEV and TEST evaluation sets, on the other hand, are re-processed by professional human transcribers to ensure high transcription quality.Baseline systems are provided for popular speech recognition toolkits, namely Athena, ESPnet, Kaldi and Pika.

We propose two end-to-end neural configurations for language diarization on bilingual code-switching speech. The first, a BLSTM-E2E architecture, includes a set of stacked bidirectional LSTMs to compute embeddings and incorporates the deep clustering loss to enforce grouping of languages belonging to the same class. The second, an XSA-E2E architecture, is based on an x-vector model followed by a self-attention encoder. The former encodes frame-level features into segmentlevel embeddings while the latter considers all those embeddings to generate a sequence of segment-level language labels. We evaluated the proposed methods on the dataset obtained from the shared task B in WSTCSMC 2020 and our handcrafted simulated data from the SEAME dataset. Experimental results show that our proposed XSA-E2E architecture achieved a relative improvement of 12.1% in equal error rate and a 7.4% relative improvement on accuracy compared with the baseline algorithm in the WSTCSMC 2020 dataset. Our proposed XSA-E2E architecture achieved an accuracy of 89.84% with a baseline of 85.60% on the simulated data derived from the SEAME dataset.

We recently proposed DOVER-Lap, a method for combining overlap-aware speaker diarization system outputs. DOVER-Lap improved upon its predecessor DOVER by using a label mapping method based on globally-informed greedy search. In this paper, we analyze this label mapping in the framework of a maximum orthogonal graph partitioning problem, and present three inferences. First, we show that DOVER-Lap label mapping is exponential in the input size, which poses a challenge when combining a large number of hypotheses. We then revisit the DOVER label mapping algorithm and propose a modification which performs similar to DOVER-Lap while being computationally tractable. We also derive an approximation bound for the algorithm in terms of the maximum number of hypotheses speakers. Finally, we describe a randomized local search algorithm which provides a near-optimal $(1-ε)$-approximate solution to the problem with high probability. We empirically demonstrate the effectiveness of our methods on the AMI meeting corpus. Our code is publicly available: https://github.com/desh2608/dover-lap.

Low-resource Multilingual Neural Machine Translation (MNMT) is typically tasked with improving the translation performance on one or more language pairs with the aid of high-resource language pairs. In this paper and we propose two simple search based curricula – orderings of the multilingual training data – which help improve translation performance in conjunction with existing techniques such as fine-tuning. Additionally and we attempt to learn a curriculum for MNMT from scratch jointly with the training of the translation system using contextual multi-arm bandits. We show on the FLORES low-resource translation dataset that these learned curricula can provide better starting points for fine tuning and improve overall performance of the translation system.

Several advances have been made recently towards handling overlapping speech for speaker diarization. Since speech and natural language tasks often benefit from ensemble techniques, we propose an algorithm for combining outputs from such diarization systems through majority voting. Our method, DOVER-Lap, is inspired from the recently proposed DOVER algorithm, but is designed to handle overlapping segments in diarization outputs. We also modify the pair-wise incremental label mapping strategy used in DOVER, and propose an approximation algorithm based on weighted k-partite graph matching, which performs this mapping using a global cost tensor. We demonstrate the strength of our method by combining outputs from diverse systems - clustering-based, region proposal networks, and target-speaker voice activity detection - on AMI and LibriCSS datasets, where it consistently outperforms the single best system. Additionally, we show that DOVER-Lap can be used for late fusion in multichannel diarization, and compares favorably with early fusion methods like beamforming.

This paper describes a method for overlap-aware speaker diarization. Given an overlap detector and a speaker embedding extractor, our method performs spectral clustering of segments informed by the output of the overlap detector. This is achieved by transforming the discrete clustering problem into a convex optimization problem which is solved by eigen-decomposition. Thereafter, we discretize the solution by alternatively using singular value decomposition and a modified version of non-maximal suppression which is constrained by the output of the overlap detector. Furthermore, we detail an HMM-DNN based overlap detector which performs frame-level classification and enforces duration constraints through HMM state transitions. Our method achieves a test diarization error rate (DER) of 24.0% on the mixed-headset setting of the AMI meeting corpus, which is a relative improvement of 15.2% over a strong agglomerative hierarchical clustering baseline, and compares favorably with other overlap-aware diarization methods. Further analysis on the LibriCSS data demonstrates the effectiveness of the proposed method in high overlap conditions.

We propose a novel lazy-evaluation token-group decoding algorithm with on-the-fly composition of weighted finite-state transducers (WFSTs) for large vocabulary continuous speech recognition. In the standard on-the-fly composition decoder, a base WFST and one or more incremental WFSTs are composed during decoding, and then token passing algorithm is employed to generate the lattice on the composed search space, resulting in substantial computation overhead. To improve speed, the proposed algorithm adopts 1) a token-group method, which groups tokens with the same state in the base WFST on each frame and limits the capacity of the group and 2) a lazy-evaluation method, which does not expand a token group and its source token groups until it processes a word label during decoding. Experiments show that the proposed decoder works notably up to 3 times faster than the standard on-the-fly composition decoder.

Speaker diarization is an important pre-processing step for many speech applications, and it aims to solve the "who spoke when" problem. Although the standard diarization systems can achieve satisfactory results in various scenarios, they are composed of several independently-optimized modules and cannot deal with the overlapped speech. In this paper, we propose a novel speaker diarization method: Region Proposal Network based Speaker Diarization (RPNSD). In this method, a neural network generates overlapped speech segment proposals, and compute their speaker embeddings at the same time. Compared with standard diarization systems, RPNSD has a shorter pipeline and can handle the overlapped speech. Experimental results on three diarization datasets reveal that RPNSD achieves remarkable improvements over the state-of-the-art x-vector baseline.

We explore two adaptation approaches of deep Transformer based neural language models (LMs) for automatic speech recognition. The first approach is a pretrain-finetune framework, where we first pretrain a Transformer LM on a large-scale text corpus from scratch and then adapt it to relatively small target domains via finetuning. The second approach is a mixer of dynamically weighted models that are separately trained on source and target domains, aiming to improve simple linear interpolation with dynamic weighting. We compare the two approaches with three baselines - without adaptation, merging data, and simple interpolation - on Switchboard (SWBD) and Wall Street Journal (WSJ). Experiments show that the mixer model generally performs better than baselines and finetuning. Compared with no adaptation, finetuning and the mixer approach obtain up to relative 11.5% and 14.1% WER reductions on SWBD, respectively. The mixer model also outperforms linear interpolation and merging data. On WSJ, the mixer approach achieves a new state-of-the-art WER result.

In this paper, we investigate out-of-vocabulary (OOV) word recovery in hybrid automatic speech recognition (ASR) systems, with emphasis on dynamic vocabulary expansion for both Weight Finite State Transducer (WFST)-based decoding and word-level RNNLM rescoring. We first describe our OOV candidate generation method based on a hybrid lexical model (HLM) with phoneme-sequence constraints. Next, we introduce a framework for efficient second pass OOV recovery with a dynamically expanded vocabulary, showing that, by calibrating OOV candidates’ language model (LM) scores, it significantly improves OOV recovery and overall decoding performance compared to HLM-based first pass decoding. Finally we propose an open-vocabulary word-level recurrent neural network language model (RNNLM) re-scoring framework, making it possible to re-score ASR hypotheses containing recovered OOVs, using a single word-level RNNLM ignorant of OOVs when it was trained. By evaluating OOV recovery and overall decoding performance on Spanish/English ASR ‘tasks, we show the proposed OOV recovery pipeline has the potential of an efficient open-vocab word-based ASR decoding framework, with minimal extra computation versus a standard WFST based decoding and RNNLM rescoring pipeline.

This paper presents an efficient algorithm for n-gram language model adaptation under the minimum discrimination information (MDI) principle, where an out-of-domain language model is adapted to satisfy the constraints of marginal probabilities of the in-domain data.The challenge for MDI language model adaptation is its computational complexity.By taking advantage of the backoff structure of n-gram model and the idea of hierarchical training method, originally proposed for maximum entropy (ME) language models [1], we show that MDI adaptation can be computed in linear-time complexity to the inputs in each iteration.The complexity remains the same as ME models, although MDI is more general than ME.This makes MDI adaptation practical for large corpus and vocabulary.Experimental results confirm the scalability of our algorithm on very large datasets, while MDI adaptation gets slightly worse perplexity but better word error rate results compared to simple linear interpolation.

A cache-inspired approach is proposed for neural language models (LMs) to improve long-range dependency and better predict rare words from long contexts.This approach is a simpler alternative to attention-based pointer mechanism that enables neural LMs to reproduce words from recent history.Without using attention and mixture structure, the method only involves appending extra tokens that represent words in history to the output layer of a neural LM and modifying training supervisions accordingly.A memory-augmentation unit is introduced to learn words that are particularly likely to repeat.We experiment with both recurrent neural network-and Transformer-based LMs.Perplexity evaluation on Penn Treebank and WikiText-2 shows the proposed model outperforms both LSTM and LSTM with attention-based pointer mechanism and is more effective on rare words.N -best rescoring experiments on Switchboard indicate that it benefits both very rare and frequent words.However, it is challenging for the proposed model as well as two other models with attention-based pointer mechanism to obtain good overall WER reductions.

We present PYCHAIN, a fully parallelized PyTorch implementation of end-to-end lattice-free maximum mutual information (LF-MMI) training for the so-called chain models in the Kaldi automatic speech recognition (ASR) toolkit.Unlike other Py-Torch and Kaldi based ASR toolkits, PYCHAIN is designed to be as flexible and light-weight as possible so that it can be easily plugged into new ASR projects, or other existing PyTorchbased ASR tools, as exemplified respectively by a new project PYCHAIN-EXAMPLE, and ESPRESSO, an existing end-to-end ASR toolkit.PYCHAIN's efficiency and flexibility is demonstrated through such novel features as full GPU training on numerator/denominator graphs, and support for unequal length sequences.Experiments on the WSJ dataset show that with simple neural networks and commonly used machine learning techniques, PYCHAIN can achieve competitive results that are comparable to Kaldi and better than other end-to-end ASR systems.

Always-on spoken language interfaces, e.g. personal digital assistants, rely on a wake word to start processing spoken input. We present novel methods to train a hybrid DNN/HMM wake word detection system from partially labeled training data, and to use it in on-line applications: (i) we remove the prerequisite of frame-level alignments in the LF-MMI training algorithm, permitting the use of un-transcribed training examples that are annotated only for the presence/absence of the wake word; (ii) we show that the classical keyword/filler model must be supplemented with an explicit non-speech (silence) model for good performance; (iii) we present an FST-based decoder to perform online detection. We evaluate our methods on two real data sets, showing 50%--90% reduction in false rejection rates at pre-specified false alarm rates over the best previously published figures, and re-validate them on a third (large) data set.

We introduce a lattice determinization algorithm that can operate incrementally. That is, a word-level lattice can be generated for a partial utterance and then, once we have processed more audio, we can obtain a word-level lattice for the extended utterance without redoing all the work of lattice determinization. This is relevant for ASR decoders such as those used in Kaldi, which first generate a state-level lattice and then convert it to a word-level lattice using a determinization algorithm in a special semiring. Our incremental determinization algorithm is useful when word-level lattices are needed prior to the end of the utterance, and also reduces the latency due to determinization at the end of the utterance.

Deep neural network based speaker embeddings, such as x-vectors, have been shown to perform well in text-independent speaker recognition/verification tasks. In this paper, we use simple classifiers to investigate the contents encoded by x-vector embeddings. We probe these embeddings for information related to the speaker, channel, transcription (sentence, words, phones), and meta information about the utterance (duration and augmentation type), and compare these with the information encoded by i-vectors across a varying number of dimensions. We also study the effect of data augmentation during extractor training on the information captured by x-vectors. Experiments on the RedDots data set show that x-vectors capture spoken content and channel-related information, while performing well on speaker verification tasks.

We present Espresso, an open-source, modular, extensible end-to-end neural automatic speech recognition (ASR) toolkit based on the deep learning library PyTorch and the popular neural machine translation toolkit FAIRSEQ. ESRESSO supports distributed training across GPUs and computing nodes, and features various decoding approaches commonly employed in ASR, including look-ahead word-based language model fusion, for which a fast, parallelized decoder is implemented. Espresso achieves state-of-the-art ASR performance on the WSJ, LibriSpeech, and Switchboard data sets among other end-to-end systems without data augmentation, and is 4-11x faster for decoding than similar systems (e.g. ESPNET).

Existing acoustic models can be transferred to any language with a pronunciation lexicon (lexicon) that uses the same set of sub-word units as in training. Unfortunately such lexicons are not readily available in many low-resource languages. We bypass this requirement and create lexicons by training a grapheme-to-phoneme (G2P) transducer on a subset of words from other languages for which pronunciations are available. The subset of words is selected based on how representative it is of target language text. We find that cross-language acoustic model transfer using our selection strategy outperforms selection based on language similarity, and results in ASR performance approaching that of hand-crafted rule based lexicons in the majority of cases.

This paper summarizes our acoustic modeling efforts in the Johns Hopkins University speech recognition system for the CHiME-5 challenge to recognize highly-overlapped dinner party speech recorded by multiple microphone arrays. We explore data augmentation approaches, neural network architectures, front-end speech dereverberation, beamforming and robust i-vector extraction with comparisons of our in-house implementations and publicly available tools. We finally achieved a word error rate of 69.4% on the development set, which is a 11.7% absolute improvement over the previous baseline of 81.1%, and release this improved baseline with refined techniques/tools as an advanced CHiME-5 recipe.

Recently, deep neural networks that map utterances to fixed-dimensional embeddings have emerged as the state-of-the-art in speaker recognition. Our prior work introduced x-vectors, an embedding that is very effective for both speaker recognition and diarization. This paper combines our previous work and applies it to the problem of speaker recognition on multi-speaker conversations. We measure performance on Speakers in the Wild and report what we believe are the best published error rates on this dataset. Moreover, we find that diarization substantially reduces error rate when there are multiple speakers, while maintaining excellent performance on single-speaker recordings. Finally, we introduce an easily implemented method to remove the domain-sensitive threshold typically used in the clustering stage of a diarization system. The proposed method is more robust to domain shifts, and achieves similar results to those obtained using a well-tuned threshold.

Hybrid deep neural network hidden Markov models (DNN-HMM) have achieved impressive results on large vocabulary continuous speech recognition (LVCSR) tasks. However, the recent approaches using DNN-HMM models are not explored much for text recognition. Inspired by the current work in automatic speech recognition (ASR) and machine translation, we present an open vocabulary sub-word text recognition system. The sub-word lexicon and sub-word language model (LM) helps in overcoming the challenge of recognizing out of vocabulary (OOV) words, and a time delay neural network (TDNN) and convolution neural network (CNN) based DNN-HMM optical model (OM) efficiently models the sequence dependency in the line image. We present results on 12 datasets with training data varying from 6k lines to 600k lines. The system is built for 8 languages, i.e., English, French, Arabic, Chinese, Farsi, Tamil, Russian, and Korean. We report competitive results on several commonly used handwritten and printed text datasets.

We explore training attention-based encoder-decoder ASR in low-resource settings. These models perform poorly when trained on small amounts of transcribed speech, in part because they depend on having sufficient target-side text to train the attention and decoder networks. In this paper we address this shortcoming by pretraining our network parameters using only text-based data and transcribed speech from other languages. We analyze the relative contributions of both sources of data. Across 3 test languages, our text-based approach resulted in a 20% average relative improvement over a text-based augmentation technique without pretraining. Using transcribed speech from nearby languages gives a further 20-30% relative reduction in character error rate.

Many applications of computer vision require robust systems that can parse complex structures as they evolve in time. Using a block construction task as a case study, we illustrate the main components involved in building such systems. We evaluate performance at three increasingly-detailed levels of spatial granularity on two multimodal (RGBD + IMU) datasets. On the first, designed to match the assumptions of the model, we report better than 90% accuracy at the finest level of granularity. On the second, designed to test the robustness of our model under adverse, real-world conditions, we report 67% accuracy and 91% precision at the mid-level of granularity. We show that this seemingly simple process presents many opportunities to expand the frontiers of computer vision and action recognition.

To date, the bulk of research on speaker diarization has been conducted on telephone or near-field speech. As the need for technologies capable of handling conversational speech increases, it is necessary to establish the performance of state-of-the-art systems in this domain. In this work we evaluate the performance of an ivector/PLDA-based diarization system on the AMI Meeting Corpus, comparing performance on near-field, far-field, and signal-enhanced conditions.

The lattice-free MMI objective (LF-MMI) has been used in supervised training of state-of-the-art neural network acoustic models for automatic speech recognition (ASR). With large amounts of unsupervised data available, extending this approach to the semi-supervised scenario is of significance. Finite-state transducer (FST) based supervision used with LF-MMI provides a natural way to incorporate uncertainties when dealing with unsupervised data. In this paper, we describe various extensions to standard LF-MMI training to allow the use as supervision of lattices obtained via decoding of unsupervised data. The lattices are rescored with a strong LM. We investigate different methods for splitting the lattices and incorporating frame tolerances into the supervision FST. We report results on different subsets of Fisher English, where we achieve WER recovery of 59-64% using lattice supervision, which is significantly better than using just the best path transcription.

Developing speech technologies for low-resource languages has become a very active research field over the last decade. Among others, Bayesian models have shown some promising results on artificial examples but still lack of in situ experiments. Our work applies state-of-the-art Bayesian models to unsupervised Acoustic Unit Discovery (AUD) in a real low-resource language scenario. We also show that Bayesian models can naturally integrate information from other resourceful languages by means of informative prior leading to more consistent discovered units. Finally, discovered acoustic units are used, either as the I-best sequence or as a lattice, to perform word segmentation. Word segmentation results show that this Bayesian approach clearly outperforms a Segmental-DTW baseline on the same corpus.

Self-attention - an attention mechanism where the input and output sequence lengths are the same - has recently been successfully applied to machine translation, caption generation, and phoneme recognition. In this paper we apply a restricted self-attention mechanism (with multiple heads) to speech recognition. By “restricted” we mean that the mechanism at a particular frame only sees input from a limited number of frames to the left and right. Restricting the context makes it easier to encode the position of the input - we use a I-hot encoding of the frame offset. We try introducing attention layers into TDNN architectures, and replacing LSTM layers with attention layers in TDNN+LSTM architectures. We show experiments on a number of ASR setups. We observe improvements compared to the TDNN and TDNN+LSTM baselines. Attention layers are also faster than LSTM layers in test time, since they lack recurrence.

Automatic speech recognition is more and more widely and effectively used. Nevertheless, in some automatic speech analysis tasks the state of the art is surprisingly poor. One of these is “diarization”, the task of determining who spoke when. Diarization is key to processing meeting audio and clinical interviews, extended recordings such as police body cam or child language acquisition data, and any other speech data involving multiple speakers whose voices are not cleanly separated into individual channels. Overlapping speech, environmental noise and suboptimal recording techniques make the problem harder. During the JSALT Summer Workshop at CMU in 2017, an international team of researchers worked on several aspects of this problem, including calibration of the state of the art, detection of overlaps, enhancement of noisy recordings, and classification of shorter speech segments. This paper sketches the workshop's results, and announces plans for a “Diarization Challenge” to encourage further progress.

In this paper, we use data augmentation to improve performance of deep neural network (DNN) embeddings for speaker recognition. The DNN, which is trained to discriminate between speakers, maps variable-length utterances to fixed-dimensional embeddings that we call x-vectors. Prior studies have found that embeddings leverage large-scale training datasets better than i-vectors. However, it can be challenging to collect substantial quantities of labeled data for training. We use data augmentation, consisting of added noise and reverberation, as an inexpensive method to multiply the amount of training data and improve robustness. The x-vectors are compared with i-vector baselines on Speakers in the Wild and NIST SRE 2016 Cantonese. We find that while augmentation is beneficial in the PLDA classifier, it is not helpful in the i-vector extractor. However, the x-vector DNN effectively exploits data augmentation, due to its supervised training. As a result, the x-vectors achieve superior performance on the evaluation datasets.

Lattice-rescoring is a common approach to take advantage of recurrent neural language models in ASR, where a word-lattice is generated from 1st-pass decoding and the lattice is then rescored with a neural model, and ann-gram approximation method is usually adopted to limit the search space. In this work, we describe a pruned lattice-rescoring algorithm for ASR, improving the n-gram approximation method. The pruned algorithm further limits the search space and uses heuristic search to pick better histories when expanding the lattice. Experiments show that the proposed algorithm achieves better ASR accuracies while running much faster than the standard algorithm. In particular, it brings a 4x speedup for lattice-rescoring with 4-gram approximation while giving better recognition accuracies than the standard algorithm.

In this paper we describe an extension of the Kaldi software toolkit to support neural-based language modeling, intended for use in automatic speech recognition (ASR) and related tasks. We combine the use of subword features (letter n-grams) and one-hot encoding of frequent words so that the models can handle large vocabularies containing infrequent words. We propose a new objective function that allows for training of unnormalized probabilities. An importance sampling based method is supported to speed up training when the vocabulary is large. Experimental results on five corpora show that Kaldi-RNNLM rivals other recurrent neural network language model toolkits both on performance and training speed.

We describe initial work on an extension of the Kaldi toolkit that supports weighted finite-state transducer (WFST) decoding on Graphics Processing Units (GPUs). We implement token recombination as an atomic GPU operation in order to fully parallelize the Viterbi beam search, and propose a dynamic load balancing strategy for more efficient token passing scheduling among GPU threads. We also redesign the exact lattice generation and lattice pruning algorithms for better utilization of the GPUs. Experiments on the Switchboard corpus show that the proposed method achieves identical 1-best results and lattice quality in recognition and confidence measure tasks, while running 3 to 15 times faster than the single process Kaldi decoder. The above results are reported on different GPU architectures. Additionally we obtain a 46-fold speedup with sequence parallelism and multi-process service (MPS) in GPU.

We present our work on improving the numerator graph for discriminative training using the lattice-free maximum mutual information (MMI) criterion. Specifically, we propose a scheme for creating unconstrained numerator graphs by removing time constraints from the baseline numerator graphs. This leads to much smaller graphs and therefore faster preparation of training supervisions. By testing the proposed un-constrained supervisions using factorized time-delay neural network (TDNN) models, we observe 0.5% to 2.6% relative improvement over the state-of-the-art word error rates on various large-vocabulary speech recognition databases.

In topic identification (topic ID) on real-world unstructured audio, an audio instance of variable topic shifts is first broken into sequential segments, and each segment is independently classified. We first present a general purpose method for topic ID on spoken segments in low-resource languages, using a cascade of universal acoustic modeling, translation lexicons to English, and English-language topic classification. Next, instead of classifying each segment independently, we demonstrate that exploring the contextual dependencies across sequential segments can provide large improvements. In particular, we propose an attention-based contextual model which is able to leverage the contexts in a selective manner. We test both our contextual and non-contextual models on four LORELEI languages, and on all but one our attention-based contextual model significantly outperforms the context-independent models.

Teacher-student (T-S) learning is a transfer learning approach, where a teacher network is used to “teach” a student network to make the same predictions as the teacher. Originally formulated for model compression, this approach has also been used for domain adaptation, and is particularly effective when parallel data is available in source and target domains. The standard approach uses a frame-level objective of minimizing the KL divergence between the frame-level posteriors of the teacher and student networks. However, for sequence-trained models for speech recognition, it is more appropriate to train the student to mimic the sequence-level posterior of the teacher network. In this work, we compare this sequence-level KL divergence objective with another semi-supervised sequence-training method, namely the lattice-free MMI, for unsupervised domain adaptation. We investigate the approaches in multiple scenarios including adapting from clean to noisy speech, bandwidth mismatch and channel mismatch.

Bidirectional long short-term memory (BLSTM) acoustic models provide a significant word error rate reduction compared to their unidirectional counterpart, as they model both the past and future temporal contexts. However, it is nontrivial to deploy bidirectional acoustic models for online speech recognition due to an increase in latency. In this letter, we propose the use of temporal convolution, in the form of time-delay neural network (TDNN) layers, along with unidirectional LSTM layers to limit the latency to 200 ms. This architecture has been shown to outperform the state-of-the-art low frame rate (LFR) BLSTM models. We further improve these LFR BLSTM acoustic models by operating them at higher frame rates at lower layers and show that the proposed model performs similar to these mixed frame rate BLSTMs. We present results on the Switchboard 300 h LVCSR task and the AMI LVCSR task, in the three microphone conditions.

In recent years, end-to-end approaches to automatic speech recognition have received considerable attention as they are much faster in terms of preparing resources. However, conventional multistage approaches, which rely on a pipeline of training hidden Markov models (HMM)-GMM models and tree-building steps still give the state-of-the-art results on most databases. In this study, we investigate flat-start one-stage training of neural networks using lattice-free maximum mutual information (LF-MMI) objective function with HMM for large vocabulary continuous speech recognition. We thoroughly look into different issues that arise in such a setup and propose a standalone system, which achieves word error rates (WER) comparable with that of the state-of-the-art multi-stage systems while being much faster to prepare. We propose to use full biphones to enable flat-start context-dependent (CD) modeling and show through experiments that our CD modeling approach can be almost as effective as regular tree-based CD modeling. We show that our flat-start LF-MMI setup together with this tree-free CD modeling technique achieves 10 to 25 % relative WER reduction compared to other end-to-end methods on well-known databases. The improvements are larger for smaller databases.

It is common in applications of ASR to have a large amount of data out-of-domain to the test data and a smaller amount of in-domain data similar to the test data. In this paper, we investigate different ways to utilize this out-of-domain data to improve ASR models based on Lattice-free MMI (LF-MMI). In particular, we experiment with multi-task training using a network with shared hidden layers; and we try various ways of adapting previously trained models to a new domain. Both types of methods are effective in reducing the WER versus in-domain models, with the jointly trained models generally giving more improvement.

This paper describes the JHU team's Kaldi system submission to the Arabic MGB-3: The Arabic speech recognition in the Wild Challenge for ASRU-2017. We use a weights transfer approach to adapt a neural network trained on the out-of-domain MGB-2 multi-dialect Arabic TV broadcast corpus to the MGB-3 Egyptian YouTube video corpus. The neural network has a TDNN-LSTM architecture and is trained using lattice-free maximum mutual information (LF-MMI) objective followed by sMBR discriminative training. For supervision, we fuse transcripts from 4 independent transcribers into confusion network training graphs. We also describe our own approach for speaker diarization and audio-transcript alignment. We use this to prepare lightly supervised transcriptions for training the seed system used for adaptation to MGB-3. Our primary submission to the challenge gives a multi-reference WER of 32.78% on the MGB-3 test set.

This paper investigates the application of unsupervised acoustic unit discovery for topic identification (topic ID) of spoken audio documents. The acoustic unit discovery method is based on a non-parametric Bayesian phone-loop model that segments a speech utterance into phone-like categories. The discovered phone-like (acoustic) units are further fed into the conventional topic ID framework. Using multilingual bottleneck features for the acoustic unit discovery, we show that the proposed method outperforms other systems that are based on cross-lingual phoneme recognizer.

The environmental robustness of DNN-based acoustic models can be significantly improved by using multi-condition training data. However, as data collection is a costly proposition, simulation of the desired conditions is a frequently adopted strategy. In this paper we detail a data augmentation approach for far-field ASR. We examine the impact of using simulated room impulse responses (RIRs), as real RIRs can be difficult to acquire, and also the effect of adding point-source noises. We find that the performance gap between using simulated and real RIRs can be eliminated when point-source noises are added. Further we show that the trained acoustic models not only perform well in the distant-talking scenario but also provide better results in the close-talking scenario. We evaluate our approach on several LVCSR tasks which can adequately represent both scenarios.

Acoustic unit discovery (AUD) is a process of automatically identifying a categorical acoustic unit inventory from speech and producing corresponding acoustic unit tokenizations. AUD provides an important avenue for unsupervised acoustic model training in a zero resource setting where expert-provided linguistic knowledge and transcribed speech are unavailable. Therefore, to further facilitate zero-resource AUD process, in this paper, we demonstrate acoustic feature representations can be significantly improved by (i) performing linear discriminant analysis (LDA) in an unsupervised self-trained fashion, and (ii) leveraging resources of other languages through building a multilingual bottleneck (BN) feature extractor to give effective cross-lingual generalization. Moreover, we perform comprehensive evaluations of AUD efficacy on multiple downstream speech applications, and their correlated performance suggests that AUD evaluations are feasible using different alternative language resources when only a subset of these evaluation resources can be available in typical zero resource applications.

Modern topic identification (topic ID) systems for speech use automatic speech recognition (ASR) to produce speech transcripts, and perform supervised classification on such ASR outputs. However, under resource-limited conditions, the manually transcribed speech required to develop standard ASR systems can be severely limited or unavailable. In this paper, we investigate alternative unsupervised solutions to obtaining tokenizations of speech in terms of a vocabulary of automatically discovered word-like or phoneme-like units, without depending on the supervised training of ASR systems. Moreover, using automatic phoneme-like tokenizations, we demonstrate that a convolutional neural network based framework for learning spoken document representations provides competitive performance compared to a standard bag-of-words representation, as evidenced by comprehensive topic ID evaluations on both single-label and multi-label classification tasks.

Speech recognition systems for irregularly-spelled languages like English normally require hand-written pronunciations.In this paper, we describe a system for automatically obtaining pronunciations of words for which pronunciations are not available, but for which transcribed data exists.Our method integrates information from the letter sequence and from the acoustic evidence.The novel aspect of the problem that we address is the problem of how to prune entries from such a lexicon (since, empirically, lexicons with too many entries do not tend to be good for ASR performance).Experiments on various ASR tasks show that, with the proposed framework, starting with an initial lexicon of several thousand words, we are able to learn a lexicon which performs close to a full expert lexicon in terms of WER performance on test data, and is better than lexicons built using G2P alone or with a pruning criterion based on pronunciation probability.

Handcrafted pronunciation lexicons are widely used in modern speech recognition systems. Designing a pronunciation lexicon, however, requires tremendous amount of expert knowledge and effort, which is not practical when applying speech recognition techniques to low resource languages. In this paper, we are interested in developing speech recognition systems for logographic languages with only a small expert pronunciation lexicon. An iterative framework is proposed to generate and refine the phonetic transcripts of the training data, which will then be aligned to their word-level transcripts for grapheme-to-phoneme (G2P) model training. The G2P model trained this way covers graphemes that appear in the training transcripts (most of which are usually unseen in a small expert lexicon for logographic languages), therefore is able to generate pronunciations for all the words in the transcripts. The proposed lexicon generation procedure is evaluated on Cantonese speech recognition and keyword search tasks. Experiments show that starting from an expert lexicon of only 1K words, we are able to generate a lexicon that works reasonably well when compared with an expert-crafted lexicon of 5K words.

The point process model (PPM) for keyword search (KWS) is a whole-word parametric approach that characterizes each query type by the timing of phonetic events observed during its production. In this paper, we first extend the PPM modeling framework to operate on context-dependent phonetic event patterns instead of monophone patterns considered in the past, which provides significant KWS improvements. Second, we use the context-dependent PPMs to drive a detection-based speech recognition architecture thats runs parallel word detectors covering the whole vocabulary and uses the independent detections to construct lattices that can be used for both KWS indexing and LVCSR decoding. This strategy produces significant improvements over the original PPM KWS framework and provides an encouraging first attempt at PPM-based LVCSR.

Mismatched transcriptions of speech in a target language refers to transcriptions provided by people unfamiliar with the language, using English letter sequences. In this work, we demonstrate the value of such transcriptions in building an ASR system for the target language. For different languages, we use less than an hour of mismatched transcriptions to successfully adapt baseline multilingual models built with no access to native transcriptions in the target language. The adapted models provide up to 25% relative improvement in phone error rates on an unseen evaluation set.

In this paper, we extend the deep long short-term memory (DL-STM) recurrent neural networks by introducing gated direct connections between memory cells in adjacent layers. These direct links, called highway connections, enable unimpeded information flow across different layers and thus alleviate the gradient vanishing problem when building deeper LSTMs. We further introduce the latency-controlled bidirectional LSTMs (BLSTMs) which can exploit the whole history while keeping the latency under control. Efficient algorithms are proposed to train these novel networks using both frame and sequence discriminative criteria. Experiments on the AMI distant speech recognition (DSR) task indicate that we can train deeper LSTMs and achieve better improvement from sequence training with highway LSTMs (HLSTMs). Our novel model obtains 43.9/47.7% WER on AMI (SDM) dev and eval sets, outperforming all previous works. It beats the strong DNN and DLSTM baselines with 15.7% and 5.3% relative improvement respectively.

Robotic surgery and other minimally-invasive surgical techniques are an integral part of patient care, and readily yield large amounts of data. Surgical tool motion (kinematic data) contains information that is useful for assessment and education. Typically, assessment and education tools that rely upon the kinematic data require substantial manual processing such as activity annotations. The goal of this paper was to develop an automated method to align surgical recordings and assign activity annotations. We developed an approach based on unsupervised alignment to efficient annotate kinematic data for its constituent activity segments. Our method includes extracting non-linear features from the kinematic data using a stacked de-noising autoencoder, and using modified dynamic time warping to align the kinematic data from different trials of the study task. We combined alignment between a test and one or a small set of template trials (with prior manual annotations) with voting based on kernel density estimation to transfer labels from the template to the test trial. Our experiments on performance of this method using two datasets captured in the training laboratory demonstrate an accuracy of 72% to 94% for annotating activity segments within a surgical training task. Our findings are robust to data captured from several surgeons, and to deviations in activity from a canonical activity sequence.

In this study, we investigate an end-to-end text-independent speaker verification system. The architecture consists of a deep neural network that takes a variable length speech segment and maps it to a speaker embedding. The objective function separates same-speaker and different-speaker pairs, and is reused during verification. Similar systems have recently shown promise for text-dependent verification, but we believe that this is unexplored for the text-independent task. We show that given a large number of training speakers, the proposed system outperforms an i-vector baseline in equal error-rate (EER) and at low miss rates. Relative to the baseline, the end-to-end system reduces EER by 13% average and 29% pooled across test conditions. The fused system achieves a reduction of 32% average and 38% pooled.

Multi-style training, using data which emulates a variety of possible test scenarios, is a popular approach towards robust acoustic modeling. However acoustic models capable of exploiting large amounts of training data in a comparatively short amount of training time are essential. In this paper we tackle the problem of reverberant speech recognition using 5500 hours of simulated reverberant data. We use time-delay neural network (TDNN) architecture, which is capable of tackling long-term interactions between speech and corrupting sources in reverberant environments. By sub-sampling the outputs at TDNN layers across time steps, training time is substantially reduced. Combining this with distributed-optimization we show that the TDNN can be trained in 3 days using up to 32 GPUs. Further, iVectors are used as an input to the neural network to perform instantaneous speaker and environment adaptation. Finally, recurrent neural network language models are applied to the lattices to further improve the performance. Our system is shown to provide state-of-the-art results in the IARPA ASpIRE challenge, with 26.5% WER on the dev Jest set.

Speech translation is conventionally carried out by cascading an automatic speech recognition (ASR) and a statistical machine translation (SMT) system.The hypotheses chosen for translation are based on the ASR system's acoustic and language model scores, and typically optimized for word error rate, ignoring the intended downstream use: automatic translation.In this paper, we present a coarseto-fine model that uses features from the ASR and SMT systems to optimize this coupling.We demonstrate that several standard features utilized by ASR and SMT systems can be used in such a model at the speech-translation interface, and we provide empirical results on the Fisher Spanish-English speech translation corpus.

A group of junior and senior researchers gathered as a part of the 2014 Frederick Jelinek Memorial Workshop in Prague to address the problem of predicting the accuracy of a nonlinear Deep Neural Network probability estimator for unknown data in a different application domain from the domain in which the estimator was trained. The paper describes the problem and summarizes approaches that were taken by the group1.

This paper introduces a new corpus of read English speech, suitable for training and evaluating speech recognition systems. The LibriSpeech corpus is derived from audiobooks that are part of the LibriVox project, and contains 1000 hours of speech sampled at 16 kHz. We have made the corpus freely available for download, along with separately prepared language-model training data and pre-built language models. We show that acoustic models trained on LibriSpeech give lower error rate on the Wall Street Journal (WSJ) test sets than models trained on WSJ itself. We are also releasing Kaldi scripts that make it easy to build these systems.

In this paper we evaluate the WER improvement from modeling pronunciation probabilities and word-specific silence probabilities in speech recognition. We do this in the context of Finite State Transducer (FST)-based decoding, where pronunciation and silence probabilities are encoded in the lexicon (L) transducer. We describe a novel way to model word-dependent silence probabilities, where in addition to modeling the probability of silence following each individual word, we also model the probability of each word appearing after silence. All of these probabilities are estimated from aligned training data, with suitable smoothing. We conduct our experiments on four commonly used automatic speech recognition datasets, namely Wall Street Journal, Switchboard, TED-LIUM, and Librispeech. The improvement from modeling pronunciation and silence probabilities is small but fairly consistent across datasets.

Data augmentation is a common strategy adopted to increase the quantity of training data, avoid overfitting and improve robustness of the models. In this paper, we investigate audio-level speech augmentation methods which directly process the raw signal. The method we particularly recommend is to change the speed of the audio signal, producing 3 versions of the original signal with speed factors of 0.9, 1.0 and 1.1. The proposed technique has a low implementation cost, making it easy to adopt. We present results on 4 different LVCSR tasks with training data ranging from 100 hours to 1000 hours, to examine the effectiveness of audio augmentation in a variety of data scenarios. An average relative improvement of 4.3% was observed across the 4 tasks.

Maximum Mutual Information (MMI) is a popular discriminative criterion that has been used in supervised training of acoustic models for automatic speech recognition. However, standard discriminative training is very sensitive to the accuracy of the transcription and hence its implementation in a semisupervised setting requires extensive filtering of data. We will show that if the supervision transcripts are not known, the natural analogue of MMI is to minimize the conditional entropy of the lattice of possible transcripts of the data. This is equivalent to the weighted average of MMI criterion over different reference transcripts, taking those reference transcripts and their weighting from the lattice itself. In this paper we describe experiments where we applied this method to the semi-supervised training of Deep Neural Network acoustic models. In our experimental setup, the proposed method gives up to 0.5% absolute WER improvement over a DNN trained with sMBR only on the transcribed part of the data. This is 37% of the improvement that we would get from doing sMBR training if we had the transcripts for the untranscribed part of the data.

In reverberant environments there are long term interactions between speech and corrupting sources. In this paper a time delay neural network (TDNN) architecture, capable of learning long term temporal relationships and translation invariant representations, is used for reverberation robust acoustic modeling. Further, iVectors are used as an input to the neural network to perform instantaneous speaker and environment adaptation, providing 10% relative improvement in word error rate. By subsampling the outputs at TDNN layers across time steps, training time is reduced. Using a parallel training algorithm we show that the TDNN can be trained on ∼ 5500 hours of speech data in 3 days using up to 32 GPUs. The TDNN is shown to provide results competitive with state of the art systems in the IARPA ASpIRE challenge, with 27.7% WER on the dev test set.

Recurrent neural network architectures have been shown to efficiently model long term temporal dependencies between acoustic events. However the training time of recurrent networks is higher than feedforward networks due to the sequential nature of the learning algorithm. In this paper we propose a time delay neural network architecture which models long term temporal dependencies with training times comparable to standard feed-forward DNNs. The network uses sub-sampling to reduce computation during training. On the Switchboard task we show a relative improvement of 6% over the baseline DNN model. We present results on several LVCSR tasks with training data ranging from 3 to 1800 hours to show the effectiveness of the TDNN architecture in learning wider temporal dependencies in both small and large data scenarios.

Modern speech recognition systems typically cluster triphone phonetic contexts using decision trees. In this paper we describe a way to build multiple complementary decision trees from the same data, for the purpose of system combination. We do this by jointly building the decision trees using an objective function that has an added entropy term to encourage diversity among the decision trees. After the trees are built, the systems are built in the standard way and the emission probabilities are combined during decoding. Experiments on multiple datasets show gains from the use of multiple trees, at the expense of evaluating multiple models in test time.

A common way to improve the performance of deep learning is to train an ensemble of neural networks and combine them during decoding. However, this is computationally expensive in test time. In this paper, we propose an diversity-penalizing ensemble training (DPET) procedure, which trains an ensemble of DNNs, whose parameters were differently initialized, and penalizes differences between each individual DNN’s output and their average output. This way each model learns to emulate the average of the whole ensemble of models, and in test time we can use one arbitrarily chosen member of the ensemble. Experimental results on a variety of speech recognition tasks show that this technique is effective, and gives us most of the WER improvement of the ensemble method while being no more expensive in test time than using a single model.

We propose an online learning algorithm based on tensor-space models. A tensor-space model represents data in a compact way, and via rank-1 approximation the weight tensor can be made highly struc-tured, resulting in a significantly smaller number of free parameters to be estimated than in comparable vector-space models. This regularizes the model complexity and makes the tensor model highly effective in situations where a large feature set is de-fined but very limited resources are avail-able for training. We apply with the pro-posed algorithm to a parsing task, and show that even with very little training data the learning algorithm based on a ten-sor model performs well, and gives signif-icantly better results than standard learn-ing algorithms based on traditional vector-space models. 1

We aim to improve spoken term detection performance by incorporating contextual information beyond traditional Ngram language models.Instead of taking a broad view of topic context in spoken documents, variability of word co-occurrence statistics across corpora leads us to focus instead the on phenomenon of word repetition within single documents.We show that given the detection of one instance of a term we are more likely to find additional instances of that term in the same document.We leverage this burstiness of keywords by taking the most confident keyword hypothesis in each document and interpolating with lower scoring hits.We then develop a principled approach to select interpolation weights using only the ASR training data.Using this re-weighting approach we demonstrate consistent improvement in the term detection performance across all five languages in the BABEL program.

In this paper we present an algorithm that produces pitch and probability-of-voicing estimates for use as features in automatic speech recognition systems. These features give large performance improvements on tonal languages for ASR systems, and even substantial improvements for non-tonal languages. Our method, which we are calling the Kaldi pitch tracker (because we are adding it to the Kaldi ASR toolkit), is a highly modified version of the getf0 (RAPT) algorithm. Unlike the original getf0 we do not make a hard decision whether any given frame is voiced or unvoiced; instead, we assign a pitch even to unvoiced frames while constraining the pitch trajectory to be continuous. Our algorithm also produces a quantity that can be used as a probability of voicing measure; it is based on the normalized autocorrelation measure that our pitch extractor uses. We present results on data from various languages in the BABEL project, and show a large improvement over systems without tonal features and systems where pitch and POV information was obtained from SAcC or getf0.

We report insights from translating Spanish conversational telephone speech into English text by cascading an automatic speech recognition (ASR) system with a statistical machine translation (SMT) system. The key new insight is that the informal register of conversational speech is a greater challenge for ASR than for SMT: the BLEU score for translating the reference transcript is 64%, but drops to 32% for translating automatic transcripts, whose word error rate (WER) is 40%. Several strategies are examined to mitigate the impact of ASR errors on the SMT output: (i) providing the ASR lattice, instead of the 1-best output, as input to the SMT system, (ii) training the SMT system on Spanish ASR output paired with English text, instead of Spanish reference transcripts, and (iii) improving the core ASR system. Each leads to consistent and complementary improvements in the SMT output. Compared to translating the 1-best output of an ASR system with 40% WER using an SMT system trained on Spanish reference transcripts, translating the output lattice of a better ASR system with 35% WER using an SMT system trained on ASR output improves BLEU from 32% to 38%.

We consider the task of identifying topics in recorded speech across many languages. We identify a statistically discriminative set of topic keywords, and examine the relationship between overall word error rate (WER), keyword-specific detection performance, and topic identification (Topic ID) performance on the Fisher Spanish corpus. Building increasingly constrained systems - from copious to limited training LVCSR to limited-vocabulary keyword spotting - we show that neither high WER (>60%) nor low-precision term detection (<;40%) are necessarily impediments to Topic ID. By using deep neural net acoustic models for keyword spotting, we can double recall and ranked retrieval performance over comparable PLP-based models and achieve Topic ID performance on par with well-trained LVCSR or human transcripts.

Recently, maxout networks have brought significant improvements to various speech recognition and computer vision tasks. In this paper we introduce two new types of generalized maxout units, which we call p-norm and soft-maxout. We investigate their performance in Large Vocabulary Continuous Speech Recognition (LVCSR) tasks in various languages with 10 hours and 60 hours of data, and find that the p-norm generalization of maxout consistently performs well. Because, in our training setup, we sometimes see instability during training when training unbounded-output nonlinearities such as these, we also present a method to control that instability. This is the “normalization layer”, which is a nonlinearity that scales down all dimensions of its input in order to stop the average squared output from exceeding one. The performance of our proposed nonlinearities are compared with maxout, rectified linear units (ReLU), tanh units, and also with a discriminatively trained SGMM/HMM system, and our p-norm units with p equal to 2 are found to perform best.

The point process model (PPM) for keyword search is a wholeword parametric modeling framework based on the timing of phonetic events rather than the evolution of frame-level phonetic likelihoods. Recent progress in PPM training and decoding algorithms has yielded state-of-the-art phonetic search performance in high-resource settings, both in terms of accuracy and computational efficiency. In this paper, we consider PPM application to low-resource settings where the amount of transcribed speech is severely limited and the pronunciation dictionary is incomplete. By using (i) state-of-the-art deep neural network acoustic models to generate phonetic events and (ii) grapheme-to-phoneme conversion to generate pronunciations for out-of-vocabulary (OOV) keywords, we find the PPM system reaches state-of-the-art OOV search performance at a small computational cost. Moreover, due to their complementary methodologies, combining PPM outputs with the LVCSR baseline produces average relative ATWV improvements of 7% and 50% for in-vocabulary and OOV keywords, respectively (16% overall).

Provides an overview of a speech-to-text (STT) and keyword search (KWS) system architecture build primarily on the top of the Kaldi toolkit and expands on a few highlights. The system was developed as a part of the research efforts of the Radical team while participating in the IARPA Babel program. Our aim was to develop a general system pipeline which could be easily and rapidly deployed in any language, independently on the language script and phonological and linguistic features of the language.

We aim to improve term detection performance by augmenting traditional N-gram language models with multiple levels of topic context. We demonstrate that incorporating complementary aspects of topicality leads to significant improvements in term detection accuracy. We represent broad topic context through document-specific latent topics inferred via a Bayesian topic model. We capture local topic context with a cache-based adaptive language model. Measured on four languages from from the IARPA Babel program, interpolating unigrams from the broad topic context improves term detection performance by up to 1% absolute via lattice re-scoring. Re-decoding with the same document-specific model improves accuracy by up to 2.1%. Adding local context via cached N-grams improves performance by up to 1.6%. A combined approach, re-decoding with latent topic information then re-scoring with the local cached N-grams gives an overall improvement of up to 2.4%. For all languages, combining broad and local topic information outperforms any individual method.

We propose a simple but effective weighted finite state transducer (WFST) based framework for handling out-of-vocabulary (OOV) keywords in a speech search task. State-of-the-art large vocabulary continuous speech recognition (LVCSR) and keyword search (KWS) systems are developed for conversational telephone speech in Tagalog. Word-based and phone-based indexes are created from word lattices, the latter by using the LVCSR system's pronunciation lexicon. Pronunciations of OOV keywords are hypothesized via a standard grapheme-to-phoneme method. In-vocabulary proxies (word or phone sequences) are generated for each OOV keyword using WFST techniques that permit incorporation of a phone confusion matrix. Empirical results when searching for the Babel/NIST evaluation keywords in the Babel 10 hour development-test speech collection show that (i) searching for word proxies in the word index significantly outperforms searching for phonetic representations of OOV words in a phone index, and (ii) while phone confusion information yields minor improvement when searching a phone index, it yields up to 40% improvement in actual term weighted value when searching a word index with word proxies.

This paper quantifies the value of pronunciation lexicons in large vocabulary continuous speech recognition (LVCSR) systems that support keyword search (KWS) in low resource languages. State-of-the-art LVCSR and KWS systems are developed for conversational telephone speech in Tagalog, and the baseline lexicon is augmented via three different grapheme-to-phoneme models that yield increasing coverage of a large Tagalog word-list. It is demonstrated that while the increased lexical coverage - or reduced out-of-vocabulary (OOV) rate - leads to only modest (ca 1%-4%) improvements in word error rate, the concomitant improvements in actual term weighted value are as much as 60%. It is also shown that incorporating the augmented lexicons into the LVCSR system before indexing speech is superior to using them post facto, e.g., for approximate phonetic matching of OOV keywords in pre-indexed lattices. These results underscore the disproportionate importance of automatic lexicon augmentation for KWS in morphologically rich languages, and advocate for using them early in the LVCSR stage.

We summarize the accomplishments of a multi-disciplinary workshop exploring the computational and scientific issues surrounding zero resource (unsupervised) speech technologies and related models of early language acquisition. Centered around the tasks of phonetic and lexical discovery, we consider unified evaluation metrics, present two new approaches for improving speaker independence in the absence of supervision, and evaluate the application of Bayesian word segmentation algorithms to automatic subword unit tokenizations. Finally, we present two strategies for integrating zero resource techniques into supervised settings, demonstrating the potential of unsupervised methods to improve mainstream technologies.

We present our work on semi-supervised learning of discriminative language models where the negative examples for sentences in a text corpus are generated using confusion models for Turkish at various granularities, specifically, word, sub-word, syllable and phone levels. We experiment with different language models and various sampling strategies to select competing hypotheses for training with a variant of the perceptron algorithm. We find that morph-based confusion models with a sample selection strategy aiming to match the error distribution of the baseline ASR system gives the best performance. We also observe that substituting half of the supervised training examples with those obtained in a semi-supervised manner gives similar results.

This paper investigates semi-supervised methods for discriminative language modeling, whereby n-best lists are “hallucinated” for given reference text and are then used for training n-gram language models using the perceptron algorithm. We perform controlled experiments on a very strong baseline English CTS system, comparing three methods for simulating ASR output, and compare the results with training with “real” n-best list output from the baseline recognizer. We find that methods based on extracting phrasal cohorts - similar to methods from machine translation for extracting phrase tables - yielded the largest gains of our three methods, achieving over half of the WER reduction of the fully supervised methods.

Discriminative language modeling is a structured classification problem. Log-linear models have been previously used to address this problem. In this paper, the standard dot-product feature representation used in log-linear models is replaced by a non-linear function parameterized by a neural network. Embeddings are learned for each word and features are extracted automatically through the use of convolutional layers. Experimental results show that as a stand-alone model the continuous space model yields significantly lower word error rate (1% absolute), while having a much more compact parameterization (60%-90% smaller). If the baseline scores are combined, our approach performs equally well.

We present a semi-supervised language modeling technique to improve search performance on words without training data. Probabilities estimated from automatic transcripts of a large corpus of in-domain audio are added to an existing LM. Requiring only pronunciations and indomain audio, our method achieves 70% of the possible gain without any supervised data or development data. This method is contrasted with techniques using additional resources. If human effort is available, we also present a transcription regime to efficiently close the gap.

The structured language model (SLM) of [1] was one of the first to successfully integrate syntactic structure into language models. We extend the SLM framework in two new directions. First, we propose a new syntactic hierarchical interpolation that improves over previous approaches. Second, we develop a general information-theoretic algorithm for pruning the underlying Jelinek-Mercer interpolated LM used in [1], which substantially reduces the size of the LM, enabling us to train on large data. When combined with hill-climbing [2] the SLM is an accurate model, space-efficient and fast for rescoring large speech lattices. Experimental results on broadcast news demonstrate that the SLM outperforms a large 4-gram LM. 1.

Inverse Document Frequency (IDF) is an important quantity in many applications, including Information Retrieval. IDF is defined in terms of document frequency, df (w), the number of documents that mention w at least once. This quantity is relatively easy to compute over textual documents, but spoken documents are more challenging. This paper considers two baselines: (1) an estimate based on the 1-best ASR output and (2) an estimate based on expected term frequencies computed from the lattice. We improve over these baselines by taking advantage of repetition. Whatever the document is about is likely to be repeated, unlike ASR errors, which tend to be more random (Poisson). In addition, we find it helpful to consider an ensemble of language models. There is an opportunity for the ensemble to reduce noise, assuming that the errors across language models are relatively uncorrelated. The opportunity for improvement is larger when WER is high. This paper considers a pairing task application that could benefit from improved estimates of df. The pairing task inputs conversational sides from the English Fisher corpus and outputs estimates of which sides were from the same conversation. Better estimates of df lead to better performance on this task.

Long-span language models, such as those involving syntactic dependencies, produce more coherent text than their n-gram counterparts. However, evaluating the large number of sentence-hypotheses in a packed representation such as an ASR lattice is intractable under such long-span models both during decoding and discriminative training. The accepted compromise is to rescore only the N-best hypotheses in the lattice using the long-span LM. We present discriminative hill climbing, an efficient and effective discriminative training procedure for long-span LMs based on a hill climbing rescoring algorithm [1]. We empirically demonstrate significant computational savings as well as error-rate reduction over N-best training methods in a state of the art ASR system for Broadcast News transcription.

Accurate estimates of language model parameters are critical for building quality text generation systems, such as automatic speech recognition. However, text training data for a domain of interest is often unavailable. Instead, we use semi-supervised model adaptation; parameters are estimated using both unlabeled in-domain data (raw speech audio) and labeled out of domain data (text.) In this work, we present a new semi-supervised language model adaptation procedure for Maximum Entropy models with n-gram features. We augment the conventional maximum likelihood training criterion on out-of-domain text data with an additional term to minimize conditional entropy on in-domain audio. Additionally, we demonstrate how to compute conditional entropy efficiently on speech lattices using first- and second-order expectation semirings. We demonstrate improvements in terms of word error rate over other adaptation techniques when adapting a maximum entropy language model from broadcast news to MIT lectures.

We address the memory problem of maximum entropy language models (MELM) with very large feature sets. Randomized techniques are employed to remove all large, exact data structures in MELM implementations. To avoid the dictionary structure that maps each feature to its corresponding weight, the feature hashing trick [1] [2] can be used. We also replace the explicit storage of features with a Bloom filter. We show with extensive experiments that false positive errors of Bloom filters and random hash collisions do not degrade model performance. Both perplexity and WER improvements are demonstrated by building MELM that would otherwise be prohibitively large to estimate or store.

Long-span language models that capture syntax and semantics are seldom used in the first pass of large vocabulary continuous speech recognition systems due to the prohibitive search-space of sentence-hypotheses. Instead, an N-best list of hypotheses is created using tractable n-gram models, and rescored using the long-span models. It is shown in this paper that computationally tractable variational approximations of the long-span models are a better choice than standard n-gram models for first pass decoding. They not only result in a better first pass output, but also produce a lattice with a lower oracle word error rate, and rescoring the N-best list from such lattices with the long-span models requires a smaller N to attain the same accuracy. Empirical results on the WSJ, MIT Lectures, NIST 2007 Meeting Recognition and NIST 2001 Conversational Telephone Recognition data sets are presented to support these claims.

We present several modifications of the original recurrent neural network language model (RNN LM).While this model has been shown to significantly outperform many competitive language modeling techniques in terms of accuracy, the remaining problem is the computational complexity. In this work, we show approaches that lead to more than 15 times speedup for both training and testing phases. Next, we show importance of using a backpropagation through time algorithm. An empirical comparison with feedforward networks is also provided. In the end, we discuss possibilities how to reduce the amount of parameters in the model. The resulting RNN model can thus be smaller, faster both during training and testing, and more accurate than the basic one.

We describe a new approach for rescoring speech lattices - with long-span language models or wide-context acoustic models - that does not entail computationally intensive lattice expansion or limited rescoring of only an N-best list. We view the set of word-sequences in a lattice as a discrete space equipped with the edit-distance metric, and develop a hill climbing technique to start with, say, the 1-best hypothesis under the lattice-generating model(s) and iteratively search a local neighborhood for the highest-scoring hypothesis under the rescoring model(s); such neighborhoods are efficiently constructed via finite state techniques. We demonstrate empirically that to achieve the same reduction in error rate using a better estimated, higher order language model, our technique evaluates fewer utterance-length hypotheses than conventional N-best rescoring by two orders of magnitude. For the same number of hypotheses evaluated, our technique results in a significantly lower error rate.

We present learning and inference algorithms for a versatile class of partially observed vector autoregressive (VAR) models for multivariate time-series data. VAR models can capture wide variety of temporal dynamics in a continuous multidimensional signal. Given a sequence of observations to be modeled by a VAR model, it is possible to estimate its parameters in closed form by solving a least squares problem. For high dimensional observations, the state space representation of a linear system is often invoked. One advantage of doing so is that we model the dynamics of a low dimensional hidden state instead of the observations, which results in robust estimation of the dynamical system parameters. The commonly used approach is to project the high dimensional observation to the low dimensional state space using a KL transform. In this article, we propose a novel approach to automatically discover the low dimensional dynamics in a switching VAR model by imposing discriminative structure on the model parameters. We demonstrate its efficacy via significant improvements in gesture recognition accuracy over a standard hidden Markov model, which does not take the state-conditional dynamics of the observations into account, on a bench-top suturing task.

In this paper, we present a novel technique for modeling the posterior probability estimates obtained from a neural net work directly in the HMM framework using the Dirichlet Mixture Models (DMMs). Since posterior probability vectors lie on a probability simplex their distribution can be modeled using DMMs. Being in an exponential family, the parameters of DMMs can be estimated in an efficient manner. Conventional approaches like TANDEM attempt to gaussianize the posteriors by suitable transforms and model them using Gaussian Mixture Models (GMMs). This requires more number of parameters as it does not exploit the fact that the probability vectors lie on a simplex. We demonstrate through TIMIT phoneme recognition experiments that the proposed technique outperforms the conventional TANDEM approach.

Useful training data for automatic speech recognition systems of colloquial speech is usually limited to expensive in-domain transcription. Broadcast news is an appealing source of easily available data to bootstrap into a new dialect. However, some languages, like Arabic, have deep linguistic differences resulting in poor cross domain performance. If no in-domain transcripts are available, but a large amount of indomain audio is, self-training may be a suitable technique to bootstrap into the domain. In this work, we attempt to adapt Modern Standard Arabic (MSA) models to Levantine Arabic without any in-domain manual transcription. We contrast with varying amounts of in-domain transcription and show that 1) Self-training is effective with only one hour of indomain transcripts. 2) Self-training is not a suitable solution to improve strong MSA models on Levantine. 3) Two metrics that quantify model bias predict self-training success. 4) Model bias explains the failure of self-training to adapt across strong domain mismatch. Index Terms: Arabic ASR, domain adaptation, self-training 1.

We propose a new iterative algorithm to reconstruct an unknown sparse signal x from a set of projected measurements y = Φx . Unlike existing methods, which rely crucially on the near orthogonality of the sampling matrix Φ , our approach makes stepwise optimal updates even when the columns of Φ are not orthogonal. We invoke a block-wise matrix inversion formula to obtain a closed-form expression for the increase (reduction) in the L2-norm of the residue obtained by removing (adding) a single element from (to) the presumed support of x . We then use this expression to design a computationally tractable algorithm to search for the nonzero components of x . We show that compared to currently popular sparsity seeking matching pursuit algorithms, each step of the proposed algorithm is locally optimal with respect to the actual objective function. We demonstrate experimentally that the algorithm significantly outperforms conventional techniques in recovering sparse signals whose nonzero values have exponentially decaying magnitudes or are distributed N(0,1) .

Given a number of machine translations of a source segment, the goal of system combination is to produce a new translation that has better quality than all of them. This paper describes a number of improvements that were recently added to the JHU system combination scheme: (i) A hypothesis ranking technique which orders the system outputs, on a per-segment basis, according to predicted translation quality, thus improving a subsequent incremental combination step. (ii) A two-pass combination procedure, which first produces several combination outputs with the given translations, and then performs one more combination step with these new outputs. Results from the NIST MT09 informal system combination evaluation on Arabic-to-English and Urdu-to-English1show that both approaches offer significant BLEU and TER gains over a baseline JHU combination scheme.

A new recurrent neural network based language model (RNN LM) with applications to speech recognition is presented. Results indicate that it is possible to obtain around 50% reduction of perplexity by using mixture of several RNN LMs, compared to a state of the art backoff language model. Speech recognition experiments show around 18% reduction of word error rate on the Wall Street Journal task when comparing models trained on the same amount of data, and around 5% on the much harder NIST RT05 task, even when the backoff model is trained on much more data than the RNN LM. We provide ample empirical evidence to suggest that connectionist language models are superior to standard n-gram techniques, except their high computational (training) complexity. Index Terms: language modeling, recurrent neural networks, speech recognition

A novel self-supervised discriminative training method for estimating language models for automatic speech recognition (ASR) is proposed. Unlike traditional discriminative training methods that require transcribed speech, only untranscribed speech and a large text corpus is required. An exponential form is assumed for the language model, as done in maximum entropy estimation, but the model is trained from the text using a discriminative criterion that targets word confusions actually witnessed in first-pass ASR output lattices. Specifically, model parameters are estimated to maximize the likelihood ratio between words w in the text corpus and w's cohorts in the test speech, i.e. other words that w competes with in the test lattices. Empirical results are presented to demonstrate statistically significant improvements over a 4-gram language model on a large vocabulary ASR task.

Pronunciation information is available in large quantities on the Web, in the form of IPA and ad-hoc transcriptions. We describe techniques for extracting candidate pronunciations from Web pages and associating them with orthographic words, filtering out poorly extracted pronunciations, normalizing IPA pronunciations to better conform to a common transcription standard, and generating phonemic from ad-hoc transcriptions. We show improvements on a letter-to-phoneme task when using web-derived vs. Pronlex pronunciations.

The problem of content-based image and video retrieval with textual queries is often posed as that of visual concept classification, where classifiers for a set of predetermined visual concepts are trained using a set of manually annotated images. Such a formulation implicitly assumes that the training data has similar distributional characteristics as that of the data which need to be indexed. In this paper we demonstrate empirically that even within the relatively narrow domain of news videos collected from a variety of news programs and broadcasters, the assumption of distributional similarity of visual features does not hold across programs from different broadcasters. This is manifested in considerable degradation of ranked retrieval performance on novel sources. We observe that concepts whose spatial locations remain relatively fixed between various sources are also more robust to source mismatches, and vice versa. We also show that a simple averaging of multiple visual detectors is more robust than any of the individual detectors. Furthermore, we show that for certain sources using only 20% of the available annotated data can bridge roughly 80% of the performance drop, while others can require larger amounts of annotated data.

This paper addresses the adjustment of the language model (LM) scaling factor of an automatic speech recognition (ASR) system for a new domain using only un-transcribed speech. The main idea is to replace the (unavailable) reference transcript with an automatic transcript generated by an independent ASR system, and adjust parameters using this sloppy reference. It is shown that despite its fairly high error rate (ca. 35%), choosing the scaling factor to minimize disagreement with the erroneous transcripts is still an effective recipe for model selection. This effectiveness is demonstrated by adjusting an ASR system trained on Broadcast News to transcribe the MIT Lectures corpus. An ASR system for telephone speech produces the sloppy reference, and optimizing towards it yields a nearly optimal LM scaling factor for the MIT Lectures corpus. Index Terms: LVCSR, language modeling, domain adaptation, semi-supervised learning, unsupervised learning

Indexing and retrieval of speech content in various forms such as broadcast news, customer care data and on-line media has gained a lot of interest for a wide range of applications, from customer analytics to on-line media search. For most retrieval applications, the speech content is typically first converted to a lexical or phonetic representation using automatic speech recognition (ASR). The first step in searching through indexes built on these representations is the generation of pronunciations for named entities and foreign language query terms. This paper summarizes the results of the work conducted during the 2008 JHU Summer Workshop by the Multilingual Spoken Term Detection team, on mining the web for pronunciations and analyzing their impact on spoken term detection. We will first present methods to use the vast amount of pronunciation information available on the Web, in the form of IPA and ad-hoc transcriptions. We describe techniques for extracting candidate pronunciations from Web pages and associating them with orthographic words, filtering out poorly extracted pronunciations, normalizing IPA pronunciations to better conform to a common transcription standard, and generating phonemic representations from ad-hoc transcriptions. We then present an analysis of the effectiveness of using these pronunciations to represent Out-Of-Vocabulary (OOV) query terms on the performance of a spoken term detection (STD) system. We will provide comparisons of Web pronunciations against automated techniques for pronunciation generation as well as pronunciations generated by human experts. Our results cover a range of speech indexes based on lattices, confusion networks and one-best transcriptions at both word and word fragments levels.

This paper addresses the detection of OOV segments in the output of a large vocabulary continuous speech recognition (LVCSR) system. First, standard confidence measures from frame-based wordand phone- posteriors are investigated. Substantial improvement is obtained when posteriors from two systems — strongly constrained (LVCSR) and weakly constrained (phone posterior estimator) are combined. We show that this approach is also suitable for detection of general recognition errors. All results are presented on WSJ task with reduced recognition vocabulary.

Current approaches to automatic spoken language identification (LID) assume the availability of a large corpus of manually language-labeled speech samples for training statistical classifiers. We investigate two methods of active learning to significantly reduce the amount of labeled speech needed for training LID systems. Starting with a small training set, an automated method is used to select samples from a corpus of unlabeled speech, which are then labeled and added to the training pool - one selection method is based on a previously known entropy criterion, and another on a novel likelihood-ratio criterion. We demonstrate LID performance comparable to a large training corpus using only a tenth of the training data. A further 40% improvement in LID performance is obtained using a third of the training data. Finally, we show that our novel selection method is more robust to variance in the unlabeled pool than the entropy based method.

We investigate methods for unsupervised learning of sub-word acoustic units of a language directly from speech. We demonstrate that states of a hidden Markov model “grown ” using a novel modification of the maximum likelihood successive state splitting algorithm correspond very well with the phones of the language. In particular, the correspondence between the Viterbi state sequence for unseen speech from the training speaker and the phone transcription of the speech is over 85%, and generalizes to a large extent ( ∼ 63%) to speech from a different speaker. Furthermore, we are able to bridge more than half the gap between the speaker-dependent and cross-speaker correspondence of the automatically learned units to phones ( ∼ 75% accuracy) by unsupervised adaptation via MLLR. 1.

Multilingual speech processing continues to develop as speech technology spreads to heterogeneous clients and applications. We address a distinct problem of code-switching — the spontaneous but occasional use, within speech in one language (referred to as L1), of words, phrases, expressions or idioms from a second language (L2). We examine two alternatives for modeling the acoustics of such words: creation of L1 pronunciations for the out-of-language (OOL) words for use with L1 acoustic models, and retention of their L2 pronunciations for use with multilingual acoustic models. We test the hypothesis that the latter is a better acoustic model for OOL words. We develop a set of lexica in IPA form, a global phoneme inventory, and handle the problem of L2 word pronunciation by creating linguistically motivated pairwise mappings. We show that retention of L2 pronunciations with multilingual acoustic models better explains the observations when restricted to a forced alignment. Index Terms: Cross-lingual and multi-lingual processing, Automatic speech recognition, Accent and language identification, Spoken language resources and annotation 1.

Csiszar introduced the mutual information of order alpha in [1] as a parameterized version of Shannon's mutual information. It involves a minimization over the probability space, and it cannot be computed in closed form. An alternating minimization algorithm for its computation is presented in this paper, along with a proof of its convergence. Furthermore, it is proved that the algorithm is an instance of the Csiszar-Tusnady iterative minimization procedure [2].

System combination is a technique which has been shown to yield significant gains in speech recognition and machine translation. Most combination schemes perform an alignment between different system outputs in order to produce lattices (or confusion networks), from which a composite hypothesis is chosen, possibly with the help of a large language model. The benefit of this approach is two-fold: (i) whenever many systems agree with each other on a set of words, the combination output contains these words with high confidence; and (ii) whenever the systems disagree, the language model resolves the ambiguity based on the (probably correct) agreed upon context. The case of machine translation system combination is more challenging because of the different word orders of the translations: the alignment has to incorporate computationally expensive movements of word blocks. In this paper, we show how one can combine translation outputs efficiently, extending the incremental alignment procedure of (A-V.I. Rosti et al., 2008). A comparison between different system combination design choices is performed on an Arabic speech translation task.

Iterative denoising trees were used by Karakos et al. (2005) for unsupervised hierarchical clustering. The tree construction involves projecting the data onto low-dimensional spaces, as a means of smoothing their empirical distributions, as well as splitting each node based on an information-theoretic maximization objective. In this paper, we improve upon the work of (Karakos et al., 2005) in two ways: (i) the amount of computation spent searching for a good projection at each node now adapts to the intrinsic dimensionality of the data observed at that node; (ii) the objective at each node is to find a split which maximizes a generalized form of mutual information, the Jensen-Renyi divergence; this is followed by an iterative Naive Bayes classification. The single parameter α of the Jensen-Renyi divergence is chosen based on the "strapping" methodology, which learns a meta-classifier on a related task. Compared with the sequential information bottleneck method, our procedure produces state-of-the-art results on an unsupervised categorization task of documents from the "20 Newsgroups" dataset.

The random forest language model (RFLM) has shown encouraging results in several automatic speech recognition (ASR) tasks but has been hindered by practical limitations, notably the space-complexity of RFLM estimation from large amounts of data. This paper addresses large-scale training and testing of the RFLM via an efficient disk-swapping strategy that exploits the recursive structure of a binary decision tree and the local access property of the tree-growing algorithm, redeeming the full potential of the RFLM, and opening avenues of further research, including useful comparisons with n-gram models. Benefits of this strategy are demonstrated by perplexity reduction and lattice rescoring experiments using a state-of-the-art ASR system. Index Terms: random forest language model, large-scale training, data scaling, speech recognition

The maximum likelihood set (MLS) is a novel candidate for nonparametric probability estimation from small samples that permits incorporating prior or structural knowledge into the estimator. It is a set of probability distributions which assign to the observed type (or empirical distribution) a likelihood that is no lower than the likelihood they assign to any other type. The MLS has been shown to have many highly desirable properties, including strong consistency of MLS-based estimates; yet the probability that the MLS contains the data-generating distribution may be arbitrarily small. In this paper, we propose to overcome this shortcoming via an epsiv-fattening of the MLS. The proposed set, called the High Likelihood Set (HLS), with epsiv rarr 0 slowly in sample size, ensures that the HLS contains the data- generating distribution with arbitrarily large probability, while retaining most desirable properties of the MLS. In particular, the HLS provides a "high-probability" bound on the estimation error, and experimental results in statistical language modeling show improved operational performance from HLS-based estimates.

Adaptation of hidden Markov model (HMM) parameters to individual speakers is known to provide considerable improvements over speaker-independent speech recognition systems. This paper applies this idea of model adaptation to a content-based video retrieval system that uses HMMs, with different sources of video treated analogously to different speakers. Source-independent HMMs are adapted to each video-source using the maximum a posteriori probability (MAP) and maximum likelihood linear regression (MLLR) techniques. It is shown that MLLR is not effective in modeling source variability in video, while MAP is highly effective. An overall improvement of 39% is demonstrated in video retrieval performance on the TRECVID 2005 benchmark test over a competitive baseline system via source-adaptation and improved use of the HMM likelihoods in retrieval

The maximum likelihood set (MLS) was recently introduced in B. Jedynak and S. Khudanpur (2005) as an effective, parameter-free technique for estimating a probability mass function (pmf) from sparse data. The MLS contains all pmfs that assign merely a higher likelihood to the observed counts than to any other set of counts, for the same sample size. In this paper, the MLS is extended to the case of conditional pmf estimation. First, it is shown that, when the criterion for selecting a pmf from the MLS is the KL-divergence, the selected conditional pmf naturally has a back-off form, except for a ceiling on the probability of high frequency symbols that are not seen in particular contexts. Second, the pmf has a sparse parameterization, leading to efficient algorithms for KL-divergence minimization. Experimental results from bigram and trigram language modeling indicate that pmfs selected from the MLS are competitive with state-of-the-art estimates

Integrated sensing and processing decision trees (ISPDT) (Priebe et al. (2004)) were introduced as a tool for supervised classification of high-dimensional data. In this paper, we consider the problem of unsupervised classification, through a recursive construction of ISPDT, where at each internal node the data (i) are split into clusters, and (ii) are transformed independently of other clusters, guided by some optimization objective. We show that the maximization of information-theoretic quantities such as mutual information and /spl alpha/-divergences is theoretically justified for growing ISPDT, assuming that each data point is generated by a finite-memory random process given the class label. Furthermore, we present heuristics that perform the maximization in a greedy manner, and we demonstrate their effectiveness with empirical results from multispectral imaging.

In this paper we describe a novel approach for jointly modeling the text and the visual components of multimedia documents for the purpose of information retrieval(IR). We propose a novel framework where individual components are developed to model different relationships between documents and queries and then combined into a joint retrieval framework. In the state-of-the-art systems, a late combination between two independent systems, one analyzing just the text part of such documents, and the other analyzing the visual part without leveraging any knowledge acquired in the text processing, is the norm. Such systems rarely exceed the performance of any single modality (i.e. text or video) in information retrieval tasks. Our experiments indicate that allowing a rich interaction between the modalities results in significant improvement in performance over any single modality. We demonstrate these results using the TRECVID03 corpus, which comprises 120 hours of broadcast news videos. Our results demonstrate over 14 % improvement in IR performance over the best reported text-only baseline and ranks amongst the best results reported on this corpus.

This paper introduces a novel method for automatic annotation of images with keywords from a generic vocabulary of concepts or objects for the purpose of content-based image retrieval. An image, represented as sequence of feature-vectors characterizing low-level visual features such as color, texture or oriented-edges, is modeled as having been stochastically generated by a hidden Markov model, whose states represent concepts. The parameters of the model are estimated from a set of manually annotated (training) images. Each image in a large test collection is then automatically annotated with the a posteriori probability of concepts present in it. This annotation supports content-based search of the image-collection via keywords. Various aspects of model parameterization, parameter estimation, and image annotation are discussed. Empirical retrieval results are presented on two image-collections | COREL and key-frames from TRECVID. Comparisons are made with two other recently developed techniques on the same datasets.

We propose a new method for estimating the probability mass function (pmf) of a discrete and finite random variable from a small sample. We focus on the observed counts--the number of times each value appears in the sample--and define the maximum likelihood set (MLS) as the set of pmfs that put more mass on the observed counts than on any other set of counts possible for the same sample size. We characterize the MLS in detail in this article. We show that the MLS is a diamond-shaped subset of the probability simplex [0,1]k bounded by at most k x (k-1) hyper-planes, where k is the number of possible values of the random variable. The MLS always contains the empirical distribution, as well as a family of Bayesian estimators based on a Dirichlet prior, particularly the well-known Laplace estimator. We propose to select from the MLS the pmf that is closest to a fixed pmf that encodes prior knowledge. When using Kullback-Leibler distance for this selection, the optimization problem comprises finding the minimum of a convex function over a domain defined by linear inequalities, for which standard numerical procedures are available. We apply this estimate to language modeling using Zipf's law to encode prior knowledge and show that this method permits obtaining state-of-the-art results while being conceptually simpler than most competing methods.

Statistical language model estimation requires large amounts of domain-specific text, which is difficult to obtain in many languages. We propose techniques which exploit domain-specific text in a resource-rich language to adapt a language model in a resource-deficient language. A primary advantage of our technique is that in the process of cross-lingual language model adaptation, we do not rely on the availability of any machine translation capability. Instead, we assume that only a modest-sized collection of story-aligned document-pairs in the two languages is available. We use ideas from cross-lingual latent semantic analysis to develop a single low-dimensional representation shared by words and documents in both languages, which enables us to (i) find documents in the resource-rich language pertaining to a specific story in the resource-deficient language, and (ii) extract statistics from the pertinent documents to adapt a language model to the story of interest. We demonstrate significant reductions in perplexity and error rates in a Mandarin speech recognition task using this technique.

In-domain texts for estimating statistical language models are not easily found for most languages of the world. We present two techniques to take advantage of in-domain text resources in other languages. First, we extend the notion of lexical triggers, which have been used monolingually for language model adaptation, to the cross-lingual problem, permitting the construction of sharper language models for a target-language document by drawing statistics from related documents in a resource-rich language. Next, we show that cross-lingual latent semantic analysis is similarly capable of extracting useful statistics for language modeling. Neither technique requires explicit translation capabilities between the two languages! We demonstrate significant reductions in both perplexity and word error rate on a Mandarin speech recognition task by using these techniques.

The success of statistical language modeling techniques is crucially dependent on the availability of a large amount training text. For a language in which such large text collections are not available, methods have recently been proposed to take advantage of a resource-rich language, together with cross-lingual information retrieval and machine translation, to sharpen language models for the resource-deficient language. In this paper, we describe investigations into such language models for an automatic speech recognition system for Mandarin Broadcast News. By exploiting a large side-corpus of contemporaneous English news articles to adapt a static Chinese language model to the news story being transcribed, we demonstrate significant improvements in recognition accuracy. The improvement from using English text is greater when less Chinese text is available to estimate the static language model. We also compare our cross-lingual adaptation to monolingual topic-dependent language model adaptation, and achieve further gains by combining the two adaptation techniques.

No abstract available.

Searching is inherently a user-centered process; people pose the questions for which machines seek answers, and ultimately people judge the degree to which retrieved documents meet their needs. Rapid development of interactive systems that use queries expressed in one language to search documents written in another poses five key challenges: (1) interaction design, (2) query formulation, (3) cross-language search, (4) construction of translated summaries, and (5) machine translation. This article describes the design of MIRACLE, an easily extensible system based on English queries that has previously been used to search French, German, and Spanish documents, and explains how the capabilities of MIRACLE were rapidly extended to accommodate Cebuano and Hindi. Evaluation results for the cross-language search component are presented for both languages, along with results from a brief full-system interactive experiment with Hindi. The article concludes with some observations on directions for further research on interactive cross-language information retrieval.

Maximum entropy (ME) techniques have been successfully used to combine different sources of linguistically meaningful constraints in language models. However, most of the current ME models can only be used for small corpora, since the computational load in training ME models for large corpora is unbearable. This problem is especially severe when non-local dependencies are considered. In this paper, we show how to train and use topic-dependent ME models efficiently for a very large corpus, Broadcast News (BN). The training time is greatly reduced by hierarchical training and divide-and-conquer approaches. The computation in using the model is also simplified by pre-normalizing the denominators of the ME model. We report new speech recognition results showing improvement with the topic model relative to the standard N-gram model for the Broadcast News task.

We propose methods to exploit contemporary news articles in a resource rich language, together with cross-language information retrieval and machine translation, to sharpen language models for a news story in a language with fewer linguistic resources. We report experimental results on storyspecific Chinese language models that use cues from a parallel corpus of English news stories. We demonstrate that even with fairly crude cross-language information retrieval, level-1 machine translation and simple linear interpolation, a significant (18%) reduction in perplexity may be obtained over a Chinese trigram model. We also demonstrate that this method of sharpening the Chinese language model is complementary to other techniques like topic dependent modeling, and the two in combination result in an even greater reduction in perplexity (28%).

We consider the estimation of the order, i.e., the number of hidden states, of a special class of discrete-time finite-alphabet hidden Markov sources. This class can be characterized in terms of equivalent renewal processes. No a priori bound is assumed on the maximum. permissible order. An order estimator based on renewal types is constructed, and is shown to be strongly consistent by computing the precise asymptotics of the probability of estimation error. The probability of underestimation of the true order decays exponentially in the number of observations while the probability of overestimation goes to zero sufficiently fast. It is further shown that this estimator has the best possible error exponent in a large class of estimators. Our results are also valid for the general class of binary independent-renewal processes with finite mean renewal times.

The thesis concerns the development of a large vocabulary continuous speech recognition (LVCSR) system for highly inflectional languages, with special emphasis on the language modeling. An idea and usage of the automatic speech recognition is introduced and the basic principles of the statistical approach to the speech recognition and the decomposition of the system into basic components are explained. An overview of the existing statistical language modeling techniques is given and methods of inferring reliable probability estimates from sparse data and measures of the language model quality are described. There are offered a theoretical background to the finite-state machinery and the application of the finite-state machine framework to LVCSR. The goals of the thesis were to build a LVCSR system for the Czech language using standard techniques that were used for English and to analyze the system performance and propose and implement techniques that would improve the recognition accuracy. The development of the baseline system is described. The Czech language properties, especially from the automatic speech recognition point of view, were analyzed. The outcomes of this theoretical analysis are exploited and language models that take into account the specific features of the Czech language are presented. There is given a description of the class-based language models that strengthen the language model robustness and therefore reduce the perplexity and consequently improve the recognition accuracy. And finally a model that uses subword parts (morphemes) as the basic language modeling units is introduced. Such model offers a better coverage of an unknown text in comparison with standard word-based models given the same vocabulary size.

The Structured Language Model (SLM) recently introduced by Chelba and Jelinek is a powerful general formalism for exploiting syntactic dependencies in a left-to-right language model for applications such as speech and handwriting recognition, spelling correction, machine translation, etc. Unlike traditional N-gram models, optimal smoothing techniques -- discounting methods and hierarchical structures for back-off -- are still being developed for the SLM. In the SLM, the statistical dependencies of a word on immediately preceding words, preceding syntactic heads, non-terminal labels, etc., are parameterized as overlapping N-gram dependencies. Statistical dependencies in the parser and tagger used by the SLM also have N-gram like structure. Deleted interpolation has been used to combine these N-gram like models. We demonstrate on two different corpora -- WSJ and Switchboard -- that more recent modified back-off strategies and nonlinear interpolation methods considerably lower the perplexity of the SLM. Improvement in word error rate is also demonstrated on the Switchboard corpus. 1.

We describe procedures and experimental results using speech from diverse source languages to build an ASR system for a single target language. This work is intended to improve ASR in languages for which large amounts of training data are not available. We have developed both knowledge-based and automatic methods to map phonetic units from the source languages to the target language. We employed HMM adaptation techniques and discriminative model combination to combine acoustic models from the individual source languages for recognition of speech in the target language. Experiments are described in which Czech Broadcast News is transcribed using acoustic models trained from small amounts of Czech read speech augmented by English, Spanish, Russian, and Mandarin acoustic models.

It is widely acknowledged that pronunciations in spontaneous speech differ significantly from citation form. For this reason, pronunciation modeling has received considerable attention in recent automatic speech recognition literature. Most of the attention however has focused on describing an alternate pronunciation as a different sequence of phonetic units using the same inventory of phones which describe canonical pronunciations. Analysis of manual phonetic transcription of conversational speech reveals a large number (>20%) of cases of genuine ambiguity: instances where human labelers disagree on the identity of the surface form. The authors investigate and characterize the acoustic evidence in the context of this ambiguity. They show that when a pronunciation change occurs, it is often the case that neither the canonical nor the alternate phone represent the acoustics very well. Based on this analysis, two methods for accommodating pronunciation ambiguity are developed. The first method attempts to resolve the ambiguity by separately modeling each baseform/surface-form pair. The second method treats the surface form as a hidden variable and "averages out" the ambiguity.

The use of syntactic structure in general and heads of syntactic constituents in particular has recently been shown to be beneficial for statistical language modeling. The paper provides an insightful analysis of this role of syntactic structure. It is shown that the predictive power of syntactic heads is mostly complementary to the predictive power of N-grams: they help in positions where an intervening phrase or clause separates the heads from the word being predicted, making the N-gram a poor predictor. Furthermore, a significant portion of this predictive power comes in the form of a more sophisticated back-off effect via the syntactic categories (nonterminal tags) of the heads. Finally, it is shown that using the categories of the syntactic heads is better than using the categories (part-of-speech tags) of the two preceding words, confirming that it is the syntactic analysis and not just the improved back-off strategy which leads to improvements over N-gram models. Experimental results for perplexity and word error rate are presented on the Switchboard corpus to support this analysis.

Maximum entropy language modeling techniques combine different sources of statistical dependence, such as syntactic relationships, topic cohesiveness and collocation frequency, in a unified and effective language model. These techniques however are also computationally very intensive, particularly during model estimation, compared to the more prevalent alternative of interpolating several simple models, each capturing one type of dependency. In this paper we present ways which significantly reduce this complexity by reorganizing the required computations. We show that in case of a model with N-gram constraints, each iteration of the parameter estimation algorithm requires the same amount of computation as estimating a comparable back-off N-gram model. In general, the computational cost of each iteration in model estimation is linear in the number of distinct "histories" seen in the training corpus, times a model-class dependent factor. The reorganization focuses mainly on reducing this...

A new statistical language model is presented which combines collocational dependencies with two important sources of long-range statistical dependence: the syntactic structure and the topic of a sentence. These dependencies or constraints are integrated using the maximum entropy technique. Substantial improvements are demonstrated over a trigram model in both perplexity and speech recognition accuracy on the Switchboard task. A detailed analysis of the performance of this language model is provided in order to characterize the manner in which it performs better than a standard N -gram model. It is shown that topic dependencies are most useful in predicting words which are semantically related by the subject matter of the conversation. Syntactic dependencies on the other hand are found to be most helpful in positions where the best predictors of the following word are not within N -gram range due to an intervening phrase or clause. It is also shown that these two methods individually enhance an N -gram model in complementary ways and the overall improvement from their combination is nearly additive.

Conversational speech exhibits considerable pronunciation variability, which has been shown to have a detrimental effect on the accuracy of automatic speech recognition. There have been many attempts to model pronunciation variation, including the use of decision trees to generate alternate word pronunciations from phonemic baseforms. Use of pronunciation models during recognition is known to improve accuracy. This paper describes the incorporation of pronunciation models into acoustic model training in addition to recognition. Subtle difficulties in the straightforward use of alternatives to canonical pronunciations are first illustrated: it is shown that simply improving the accuracy of the phonetic transcription used for acoustic model training is of little benefit. Acoustic models trained on the most accurate phonetic transcriptions result in worse recognition than acoustic models trained on canonical baseforms. Analysis of this counterintuitive result leads to a new method of accommodating nonstandard pronunciations: rather than allowing a phoneme in the canonical pronunciation to be realized as one of a few distinct alternate phones, the hidden Markov model (HMM) states of the phoneme’s model are instead allowed to share Gaussian mixture components with the HMM states of the model(s) of the alternate realization(s). Qualitatively, this amounts to making a soft decision about which surface form is realized. Quantitatively, experiments show that this method is particularly well suited for acoustic model training for spontaneous speech: a 1.7 %(absolute) improvement in recognition accuracy on the Switchboard corpus is presented.

This paper summarizes the work of the "Rapid Speech Recognizer Adaptation" team in the workshop held at Johns Hopkins University in the summer of 1998. The project addressed the modeling of dependencies between units of speech with the goal of making more effective use of small amounts of data for speaker adaptation. A variety of methods were investigated and their effectiveness in a rapid adaptation task defined on the Switchboard conversational speech corpus is reported.

Two models of statistical dependence between the acoustic model parameters of a large vocabulary conversational speech recognition (LVCSR) system are investigated for the purpose of rapid speaker- and environment-adaptation from a very small amount of speech: (i) a Gaussian multiscale process governed by a stochastic linear dynamical system on a tree, and (ii) a simple hierarchical tree-structured prior. Both methods permit Bayesian (MAP) estimation of acoustic model parameters without parameter-tying even when no samples are available to independently estimate some parameters due to the limited amount of adaptation data. Modeling methodologies are contrasted, and comparative performance of the two on the Switchboard task is presented under identical test conditions for supervised and unsupervised adaptation with controlled amounts of adaptation speech. Both methods provide significant (1% absolute) gain in accuracy over adaptation methods that do not exploit the dependence between acoustic model parameters.

A compact language model which incorporates local dependencies in the form of N-grams and long distance dependencies through dynamic topic conditional constraints is presented. These constraints are integrated using the maximum entropy principle. Issues in assigning a topic to a test utterance are investigated. Recognition results on the Switchboard corpus are presented showing that with a very small increase in the number of model parameters, reduction in word error rate and language model perplexity are achieved over trigram models. Some analysis follows, demonstrating that the gains are even larger on content-bearing words. The results are compared with those obtained by interpolating topicindependent and topic-specific N-gram models. The framework presented here extends easily to incorporate other forms of statistical dependencies such as syntactic word-pair relationships or hierarchical topic constraints. 1. INTRODUCTION Language modeling is a crucial component of systems that c...

Conversational speech exhibits considerable pronunciation variability, which has been shown to have a detrimental effect on the accuracy of automatic speech recognition. There have been many attempts to model pronunciation variation, including the use of decision trees to generate alternate word pronunciations from phonemic baseforms. Use of pronunciation models during recognition is known to improve accuracy. This paper describes the incorporation of pronunciation models into acoustic model training in addition to recognition. Subtle difficulties in the straightforward use of alternatives to canonical pronunciations are first illustrated: it is shown that simply improving the accuracy of the phonetic transcription used for acoustic model training is of little benefit. Acoustic models trained on the most accurate phonetic transcriptions result in worse recognition than acoustic models trained on canonical baseforms. Analysis of this counterintuitive result leads to a new method of accommodating nonstandard pronunciations: rather than allowing a phoneme in the canonical pronunciation to be realized as one of a few distinct alternate phones, the hidden Markov model (HMM) states of the phoneme?s model are instead allowed to share Gaussian mixture components with the HMM states of the model(s) of the alternate realization(s). Qualitatively, this amounts to making a soft decision about which surface form is realized. Quantitatively, experiments show that this method is particularly well suited for acoustic model training for spontaneous speech: a 1.7 %(absolute) improvement in recognition accuracy on the Switchboard corpus is presented.

In this paper, a speaker adaptation method is presented which computes the speaker adapted model by a weighted sum of a set of speaker dependent models. The set of weights are estimated to maximize the likelihood of the adaptation data. Then a linguistic tree is constructed to cluster the mean vectors. The means in the same linguistic class share the same weight set, while the means in different classes use different weight set to compute the adapted model. Experiments show that with as little as 1~3 sentences a significant performance improvement is obtained. As more adaptation data is available, further improvement can be obtained.

Accurately modelling pronunciation variability in conversational speech is an important component of an automatic speech recognition system. We describe some of the projects undertaken in this direction during and after WS97, the Fifth LVCSR Summer Workshop, held at Johns Hopkins University, Baltimore, in July-August, 1997. We first illustrate a use of hand-labelled phonetic transcriptions of a portion of the Switchboard corpus, in conjunction with statistical techniques, to learn alternatives to canonical pronunciations of words. We then describe the use of these alternate pronunciations in an automatic speech recognition system. We demonstrate that the improvement in recognition performance from pronunciation modelling persists as the system is enhanced with better acoustic and language models.

The problem of speech decoding is considered in a decision theoretic framework and a modified speech decoding procedure to minimize the expected risk under a general loss function is formulated. A specific word error rate loss function is considered and an implementation in an N-best list rescoring procedure is presented. Methods for estimation of the parameters of the resulting decision rules are provided for both supervised and unsupervised training. Preliminary experiments on an LVCSR task show small but statistically significant error rate improvements.