In recent years, all-neural end-to-end approaches have obtained state-of-the-art results on several challenging automatic speech recognition (ASR) tasks. However, most existing works focus on building ASR models where train and test data are drawn from the same domain. This results in poor generalization characteristics on mismatched-domains: e.g., end-to-end models trained on short segments perform poorly when evaluated on longer utterances. In this work, we analyze the generalization properties of streaming and non-streaming recurrent neural network transducer (RNN-T) based end-to-end models in order to identify model components that negatively affect generalization performance. We propose two solutions: combining multiple regularization techniques during training, and using dynamic overlapping inference. On a long-form YouTube test set, when the non-streaming RNN-T model is trained with shorter segments of data, the proposed combination improves word error rate (WER) from 22.3% to 14.8%; when the streaming RNN-T model trained on short Search queries, the proposed techniques improve WER on the YouTube set from 67.0% to 25.3%. Finally, when trained on Librispeech, we find that dynamic overlapping inference improves WER on YouTube from 99.8% to 33.0%.

End-to-end automatic speech recognition (ASR) models, including both attention-based models and the recurrent neural network transducer (RNN-T), have shown superior performance compared to conventional systems [1], [2]. However, previous studies have focused primarily on short utterances that typically last for just a few seconds or, at most, a few tens of seconds. Whether such architectures are practical on long utterances that last from minutes to hours remains an open question. In this paper, we both investigate and improve the performance of end-to-end models on long-form transcription. We first present an empirical comparison of different end-to-end models on a real world long-form task and demonstrate that the RNN-T model is much more robust than attention-based systems in this regime. We next explore two improvements to attention-based systems that significantly improve its performance: restricting the attention to be monotonic, and applying a novel decoding algorithm that breaks long utterances into shorter overlapping segments. Combining these two improvements, we show that attention-based end-to-end models can be very competitive to RNN-T on long-form speech recognition.

In conventional speech recognition, phoneme-based models outperform grapheme-based models for non-phonetic languages such as English. The performance gap between the two typically reduces as the amount of training data is increased. In this work, we examine the impact of the choice of modeling unit for attention-based encoder-decoder models. We conduct experiments on the LibriSpeech 100hr, 460hr, and 960hr tasks, using various target units (phoneme, grapheme, and word-piece); across all tasks, we find that grapheme or word-piece models consistently outperform phoneme-based models, even though they are evaluated without a lexicon or an external language model. We also investigate model complementarity: we find that we can improve WERs by up to 9% relative by rescoring N-best lists generated from a strong word-piece based baseline with either the phoneme or the grapheme model. Rescoring an N-best list generated by the phonemic system, however, provides limited improvements. Further analysis shows that the word-piece-based models produce more diverse N-best hypotheses, and thus lower oracle WERs, than phonemic models.

Maximum-likelihood estimation (MLE) is one of the most widely used approaches for training structured prediction models for textgeneration based natural language processing applications.However, besides exposure bias, models trained with MLE suffer from wrong objective problem where they are trained to maximize the word-level correct next step prediction, but are evaluated with respect to sequence-level discrete metrics such as ROUGE and BLEU.Several variants of policy-gradient methods address some of these problems by optimizing for final discrete evaluation metrics and showing improvements over MLE training for downstream tasks like text summarization and machine translation.However, policy-gradient methods suffers from high sample variance, making the training process very difficult and unstable.In this paper, we present an alternative direction towards mitigating this problem by introducing a new objective (CALCS) based on a differentiable surrogate of longest common subsequence (LCS) measure that captures sequence-level structure similarity.Experimental results on abstractive summarization and machine translation validate the effectiveness of the proposed approach.

Attention-based encoder-decoder architectures such as Listen, Attend, and Spell (LAS), subsume the acoustic, pronunciation and language model components of a traditional automatic speech recognition (ASR) system into a single neural network. In previous work, we have shown that such architectures are comparable to state-of-the-art ASR systems on dictation tasks, but it was not clear if such architectures would be practical for more challenging tasks such as voice search. In this work, we explore a variety of structural and optimization improvements to our LAS model which significantly improve performance. On the structural side, we show that word piece models can be used instead of graphemes. We also introduce a multi-head attention architecture, which offers improvements over the commonly-used single-head attention. On the optimization side, we explore synchronous training, scheduled sampling, label smoothing, and minimum word error rate optimization, which are all shown to improve accuracy. We present results with a unidirectional LSTM encoder for streaming recognition. On a 12, 500 hour voice search task, we find that the proposed changes improve the WER from 9.2% to 5.6%, while the best conventional system achieves 6.7%; on a dictation task our model achieves a WER of 4.1% compared to 5% for the conventional system.

Attention-based sequence-to-sequence models for automatic speech recognition jointly train an acoustic model, language model, and alignment mechanism. Thus, the language model component is only trained on transcribed audio-text pairs. This leads to the use of shallow fusion with an external language model at inference time. Shallow fusion refers to log-linear interpolation with a separately trained language model at each step of the beam search. In this work, we investigate the behavior of shallow fusion across a range of conditions: different types of language models, different decoding units, and different tasks. On Google Voice Search, we demonstrate that the use of shallow fusion with an neural LM with wordpieces yields a 9.1% relative word error rate reduction (WERR) over our competitive attention-based sequence-to-sequence model, obviating the need for second-pass rescoring.

Sequence-to-sequence models provide a simple and elegant solution for building speech recognition systems by folding separate components of a typical system, namely acoustic (AM), pronunciation (PM) and language (LM) models into a single neural network. In this work, we look at one such sequence-to-sequence model, namely listen, attend and spell (LAS) [1], and explore the possibility of training a single model to serve different English dialects, which simplifies the process of training multi-dialect systems without the need for separate AM, PM and LMs for each dialect. We show that simply pooling the data from all dialects into one LAS model falls behind the performance of a model fine-tuned on each dialect. We then look at incorporating dialect-specific information into the model, both by modifying the training targets by inserting the dialect symbol at the end of the original grapheme sequence and also feeding a 1-hot representation of the dialect information into all layers of the model. Experimental results on seven English dialects show that our proposed system is effective in modeling dialect variations within a single LAS model, outperforming a LAS model trained individually on each of the seven dialects by 3.1~16.5% relative.

Sequence-to-sequence models, such as attention-based models in automatic speech recognition (ASR), are typically trained to optimize the cross-entropy criterion which corresponds to improving the log-likelihood of the data. However, system performance is usually measured in terms of word error rate (WER), not log-likelihood. Traditional ASR systems benefit from discriminative sequence training which optimizes criteria such as the state-level minimum Bayes risk (sMBR) which are more closely related to WER. In the present work, we explore techniques to train attention-based models to directly minimize expected word error rate. We consider two loss functions which approximate the expected number of word errors: either by sampling from the model, or by using N-best lists of decoded hypotheses, which we find to be more effective than the sampling-based method. In experimental evaluations, we find that the proposed training procedure improves performance by up to 8.2% relative to the baseline system. This allows us to train grapheme-based, uni-directional attention-based models which match the performance of a traditional, state-of-the-art, discriminative sequence-trained system on a mobile voice-search task.

Having a sequence-to-sequence model which can operate in an online fashion is important for streaming applications such as Voice Search. Neural transducer is a streaming sequence-to-sequence model, but has shown a significant degradation in performance compared to non-streaming models such as Listen, Attend and Spell (LAS). In this paper, we present various improvements to NT. Specifically, we look at increasing the window over which NT computes attention, mainly by looking backwards in time so the model still remains online. In addition, we explore initializing a NT model from a LAS-trained model so that it is guided with a better alignment. Finally, we explore including stronger language models such as using wordpiece models, and applying an external LM during the beam search. On a Voice Search task, we find with these improvements we can get NT to match the performance of LAS.

For decades, context-dependent phonemes have been the dominant sub-word unit for conventional acoustic modeling systems. This status quo has begun to be challenged recently by end-to-end models which seek to combine acoustic, pronunciation, and language model components into a single neural network. Such systems, which typically predict graphemes or words, simplify the recognition process since they remove the need for a separate expert-curated pronunciation lexicon to map from phoneme-based units to words. However, there has been little previous work comparing phoneme-based versus grapheme-based sub-word units in the end-to-end modeling framework, to determine whether the gains from such approaches are primarily due to the new probabilistic model, or from the joint learning of the various components with grapheme-based units. In this work, we conduct detailed experiments which are aimed at quantifying the value of phoneme-based pronunciation lexica in the context of end-to-end models. We examine phoneme-based end-to-end models, which are contrasted against grapheme-based ones on a large vocabulary English Voice-search task, where we find that graphemes do indeed outperform phonemes. We also compare grapheme and phoneme-based approaches on a multi-dialect English task, which once again confirm the superiority of graphemes, greatly simplifying the system for recognizing multiple dialects.

In this paper we document our experiences with developing speech recognition for medical transcription -a system that automatically transcribes doctor-patient conversations.Towards this goal, we built a system along two different methodological lines -a Connectionist Temporal Classification (CTC) phoneme based model and a Listen Attend and Spell (LAS) grapheme based model.To train these models we used a corpus of anonymized conversations representing approximately 14,000 hours of speech.Because of noisy transcripts and alignments in the corpus, a significant amount of effort was invested in data cleaning issues.We describe a two-stage strategy we followed for segmenting the data.The data cleanup and development of a matched language model was essential to the success of the CTC based models.The LAS based models, however were found to be resilient to alignment and transcript noise and did not require the use of language models.CTC models were able to achieve a word error rate of 20.1%, and the LAS models were able to achieve 18.3%.Our analysis shows that both models perform well on important medical utterances and therefore can be practical for transcribing medical conversations.

We describe a neural network-based system for text-to-speech (TTS) synthesis that is able to generate speech audio in the voice of many different speakers, including those unseen during training. Our system consists of three independently trained components: (1) a speaker encoder network, trained on a speaker verification task using an independent dataset of noisy speech from thousands of speakers without transcripts, to generate a fixed-dimensional embedding vector from seconds of reference speech from a target speaker; (2) a sequence-to-sequence synthesis network based on Tacotron 2, which generates a mel spectrogram from text, conditioned on the speaker embedding; (3) an auto-regressive WaveNet-based vocoder that converts the mel spectrogram into a sequence of time domain waveform samples. We demonstrate that the proposed model is able to transfer the knowledge of speaker variability learned by the discriminatively-trained speaker encoder to the new task, and is able to synthesize natural speech from speakers that were not seen during training. We quantify the importance of training the speaker encoder on a large and diverse speaker set in order to obtain the best generalization performance. Finally, we show that randomly sampled speaker embeddings can be used to synthesize speech in the voice of novel speakers dissimilar from those used in training, indicating that the model has learned a high quality speaker representation.

Today's speech recognition technology is mature enough to be useful for many practical applications. In this context, it is of paramount importance to train accurate acoustic models for many languages within given resource constraints such as data, processing power, and time. Multilingual training has the potential to solve the data issue and close the performance gap between resource-rich and resource-scarce languages. Neural networks lend themselves naturally to parameter sharing across languages, and distributed implementations have made it feasible to train large networks. In this paper, we present experimental results for cross- and multi-lingual network training of eleven Romance languages on 10k hours of data in total. The average relative gains over the monolingual baselines are 4%/2% (data-scarce/data-rich languages) for cross- and 7%/2% for multi-lingual training. However, the additional gain from jointly training the languages on all data comes at an increased training time of roughly four weeks, compared to two weeks (monolingual) and one week (crosslingual).

Deep neural networks have recently become the gold standard for acoustic modeling in speech recognition systems. The key computational unit of a deep network is a linear projection followed by a point-wise non-linearity, which is typically a logistic function. In this work, we show that we can improve generalization and make training of deep networks faster and simpler by substituting the logistic units with rectified linear units. These units are linear when their input is positive and zero otherwise. In a supervised setting, we can successfully train very deep nets from random initialization on a large vocabulary speech recognition task achieving lower word error rates than using a logistic network with the same topology. Similarly in an unsupervised setting, we show how we can learn sparse features that can be useful for discriminative tasks. All our experiments are executed in a distributed environment using several hundred machines and several hundred hours of speech data.

The acoustic models in state-of-the-art speech recognition systems are based on phones in context that are represented by hidden Markov models. This modeling approach may be limited in that it is hard to incorporate long-span acoustic context. Exemplar-based approaches are an attractive alter-native, in particular if massive data and computational power are available. Yet, most of the data at Google are unsupervised and noisy. This paper investigates an exemplar-based approach under this yet not well understood data regime. A log-linear rescoring framework is used to combine the exemplar-based features on the word level with the first-pass model. This approach guarantees at least baseline performance and focuses on the refined modeling of words with sufficient data. Experimental results for the Voice Search and the YouTube tasks are presented.

The use of Deep Belief Networks (DBN) to pretrain Neural Networks has recently led to a resurgence in the use of Artificial Neural Network- Hidden Markov Model (ANN/HMM) hybrid systems for Automatic Speech Recognition (ASR). In this paper we report results of a DBN-pretrained context-dependent ANN/HMM system trained on two datasets that are much larger than any reported previously with DBN-pretrained ANN/HMM

Recent work on deep neural networks as acoustic models for automatic speech recognition (ASR) have demonstrated substantial performance improvements. We introduce a model which uses a deep recurrent auto encoder neural network to denoise input features for robust ASR. The model is trained on stereo (noisy and clean) audio features to predict clean features given noisy input. The model makes no assumptions about how noise affects the signal, nor the existence of distinct noise environments. Instead, the model can learn to model any type of distortion or additive noise given sufficient training data. We demonstrate the model is competitive with existing feature denoising approaches on the Aurora2 task, and outperforms a tandem approach where deep networks are used to predict phoneme posteriors directly. Index Terms: neural networks, robust ASR, deep learning 1.

Exemplar based recognition systems are characterized by the fact that, instead of abstracting large amounts of data into compact models, they store the observed data enriched with some annotations and infer on-the-fly from the data by finding those exemplars that resemble the input speech best. One advantage of exemplar based systems is that next to deriving what the current phone or word is, one can easily derive a wealth of meta-information concerning the chunk of audio under investigation. In this work we harvest meta-information from the set of best matching exemplars, that is thought to be relevant for the recognition such as word boundary predictions and speaker entropy. Integrating this meta-information into the recognition framework using segmental conditional random fields, reduced the WER of the exemplar based system on the WSJ Nov92 20k task from 8.2% to 7.6%. Adding the HMM-score and multiple HMM phone detectors as features further reduced the error rate to 6.6%.

This paper describes a new approach to modeling duration for LVCSR using SCARF, a toolkit for speech recognition with segmental conditional random fields. We utilize SCARF's ability to integrate long-span, segment-level features to design and test duration models that help discriminate between correct and incorrect word hypotheses. We show that the duration distributions of correct and incorrect word hypotheses differ. Given a word hypothesis in the lattice and its duration, conditional length probabilities are integrated to the SCARF system as duration features. We evaluate three kinds of duration features on Broadcast News: word, pre- and post-pausal durations, and word span confusions. Adding the duration features to SCARF results in an up to 0.3% improvement over a state of-the-art discriminatively trained baseline of 15.3% WER on a Broadcast News task.

Phoneme posterior probabilities estimated using Multi-Layer Perceptrons (MLPs) are extensively used both as acoustic scores and features for speech recognition. In this paper we explore a different application of these posteriors as phonetic event detectors for speech recognition. We show how these detectors can be built to reliably capture phonetic events in the acoustic signal by integrating both acoustic and phonetic information about sound classes. These event detectors are used along with Segmental Conditional Random Fields (SCRFs) to improve the performance of speech recognition systems on the Broadcast News task.

This paper summarizes the 2010 CLSP Summer Workshop on speech recognition at Johns Hopkins University. The key theme of the workshop was to improve on state-of-the-art speech recognition systems by using Segmental Conditional Random Fields (SCRFs) to integrate multiple types of information. This approach uses a state of-the-art baseline as a springboard from which to add a suite of novel features including ones derived from acoustic templates, deep neural net phoneme detections, duration models, modulation features, and whole word point-process models. The SCRF framework is able to appropriately weight these different information sources to produce significant gains on both die Broadcast News and Wall Street Journal tasks.

This paper addresses the problem of developing appropriate features for use in direct modeling approaches to speech recognition, such as those based on Maximum Entropy models or Segmental Conditional Random Fields. We propose a feature based on the detection of word-level templates which are discriminatively chosen based on a mutual information criterion. The templates for a word are derived directly from the MFCC feature vectors, based on self-similarity across examples. No pronunciation dictionary is used, and the resulting templates match closely to in-class examples and distantly to out-of-class examples. We utilize template detection events as input to a segmental CRF speech recognizer. We evaluate the entire scheme on a voice search task. The results show that the use of discriminative template based word detector streams improves the speech recognizer's performance over the baseline HMM results.

This paper summarizes recent work at Microsoft on the development of novel direct models. The key characteristic of our approaches is the use of long-span segment level features that relate acoustic properties directly to words. In this approach, the frame-level Markov assumption is replaced by the segment level Markov property, allowing us to extract long-span features. A key issue we address is the definition of generalizable features which allow us to model unseen words. We review two recently developed models that have this property: Flat Direct Models (FDMs), and Segmental CRFs (SCRFs). The first operates in a log-linear framework, and uses utterance level features. The second is also a log-linear model, but defines features at the word-segment level. We present new experimental results comparing the two approaches. We find that both show consistent improvements over a baseline system, and that the extra context available to the FDM enables slightly better performance in a rescoring context. This gain comes at the expense of applicability to first pass decoding, for which the SCRF is better suited.

This paper describes a new toolkit- SCARF- for doing speech recognition with segmental conditional random fields. It is designed to allow for the integration of numerous, possibly redundant segment level acoustic features, along with a complete language model, in a coherent speech recognition framework. SCARF performs a segmental analysis, where each segment corresponds to a word, thus allowing for the incorporation of acoustic features defined at the phoneme, multi-phone, syllable and word level. SCARF is designed to make it especially convenient to use acoustic detection events as input, such as the detection of energy bursts, phonemes, or other events. Language modeling is done by associating each state in the SCRF with a state in an underlying n-gram language model, and SCARF supports the joint and discriminative training of language model and acoustic model parameters. SCARF is available for download from

Information retrieval methods are frequently used for indexing and retrieving spoken documents, and more recently have been proposed for voice-search amongst a pre-defined set of business entries. In this paper, we show that these methods can be used in an even more fundamental way, as the core component in a continuous speech recognizer. Speech is initially processed and represented as a sequence of discrete symbols, specifically phoneme or multi-phone units. Recognition then operates on this sequence. The recognizer is segment-based, and the acoustic score for labeling a segment with a word is based on the TF-IDF similarity between the subword units detected in the segment, and those typically seen in association with the word. We present promising results on both a voice search task and the Wall Street Journal task. The development of this method brings us one step closer to being able to do speech recognition based on the detection of sub-word audio attributes.

This paper proposes a segmental conditional random field framework for large vocabulary continuous speech recognition. Fundamental to this approach is the use of acoustic detectors as the basic input, and the automatic construction of a versatile set of segment-level features. The detector streams operate at multiple time scales (frame, phone, multi-phone, syllable or word) and are combined at the word level in the CRF training and decoding processes. A key aspect of our approach is that features are defined at the word level, and are naturally geared to explain long span phenomena such as formant trajectories, duration, and syllable stress patterns. Generalization to unseen words is possible through the use of decomposable consistency features and our framework allows for the joint or separate discriminative training of the acoustic and language models. An initial evaluation of this framework with voice search data from the Bing Mobile (BM) application results in a 2% absolute improvement over an HMM baseline.

We introduce a direct model for speech recognition that assumes an unstructured, i.e., flat text output. The flat model allows us to model arbitrary attributes and dependences of the output. This is different from the HMMs typically used for speech recognition. This conventional modeling approach is based on sequential data and makes rigid assumptions on the dependences. HMMs have proven to be convenient and appropriate for large vocabulary continuous speech recognition. Our task under consideration, however, is the Windows Live Search for Mobile (WLS4M) task. This is a cellphone application that allows users to interact with web-based information portals. In particular, the set of valid outputs can be considered discrete and finite (although probably large, i.e., unseen events are an issue). Hence, a flat direct model lends itself to this task, making the adding of different knowledge sources and dependences straightforward and cheap. Using e.g. HMM posterior, m-gram, and spotter features, significant improvements over the conventional HMM system were observed.

A voice search system requires a speech interface that can correctly recognize spoken queries uttered by users. The recognition performance strongly relies on a robust language model. In this work, we present the use of multiple data sources, with the focus on query logs, in improving ASR language models for a voice search application. Our contributions are three folds: (1) the use of text queries from web search and mobile search in language modeling; (2) the use of web click data to predict query forms from business listing forms; and (3) the use of voice query logs in creating a positive feedback loop. Experiments show that by leveraging these resources, we can achieve recognition performance comparable to, or even better than, that of a previously deploy system where a large amount of spoken query transcripts are used in language modeling.

This paper introduces a class of discriminative features for use in maximum entropy speech recognition models. The features we propose are acoustic detectors for discriminatively determined multi-phone units. The multi-phone units are found by computing the mutual information between the phonetic subsequences that occur in the training lexicon, and the word labels. This quantity is a function of an error model governing our ability to detect phone sequences accurately (an otherwise informative sequence which cannot be reliably detected is not so useful). We show how to compute this mutual information quantity under a class of error models efficiently, in one pass over the data, for all phonetic sub-sequences in the training data. After this computation, detectors are created for a subset of highly informative units. We then define two novel classes of features based on these units: associative and transductive. Incorporating these features in a maximum entropy based direct model for Voice-Search outperforms the baseline by 24 % in sentence error rate. Index Terms: speech recognition, direct model, maximum mutual information, features

This article presents a new hypothesis alignment method for combining outputs of multiple machine translation (MT) systems. An indirect hidden Markov model (IHMM) is proposed to address the synonym matching and word ordering issues in hypothesis alignment. Unlike traditional HMMs whose parameters are trained via maximum likelihood estimation (MLE), the parameters of the IHMM are estimated indirectly from a variety of sources including word semantic similarity, word surface similarity, and a distance-based distortion penalty. The IHMM-based method significantly outperforms the state-of-the-art, TER-based alignment model in our experiments on NIST benchmark datasets. Our combined SMT system using the proposed method achieved the best Chinese-to-English translation result in the constrained training track of the 2008 NIST Open MT Evaluation.

Live search for mobile is a cellphone application that allows users to interact with Web-based information portals. Currently the implementation is focused on information related to local businesses: their phone numbers and addresses, directions, reviews, maps of the surrounding area, and traffic. This paper describes a speech-recognition interface which was recently developed for the application, which allows the users to interact by voice. The paper presents the overall architecture, the user interface, the design and implementation of the speech recognition grammars, and initial performance results indicating that for sentence level utterance recognition we achieve 60 to 65% of human capability.

This paper extends language identification (LID) techniques to a large scale accent classification task: 23-way classification of foreign-accented English. We find that a purely acoustic approach based on a combination of heteroscedastic linear discriminant analysis (HLDA) and maximum mutual information (MMI) training is very effective. In contrast to LID tasks, methods based on parallel language models prove much less effective. We focus on the Oregon Graduate Institute Foreign-Accented English dataset, and obtain a detection rate of 32%, which to our knowledge is the best reported result for 23-way accent classification.

Due to speech recognition errors, repetition can be a frequent occurrence in voice-search applications. While a proper treatment of this phenomenon requires the joint modeling of two or more utterances simultaneously, currently deployed systems typically treat the utterances independently. In this paper, we analyze the structure of repetitions and find that in at least one commercial directory assistance application, repetitions follow simple structural transformations more than 70 % of the time. We present preliminary results that suggest that significant gains are possible by explicitly modeling this structure in a joint decoding process. Index Terms: speech recognition, minimum bayes risk, joint decoding, repeated utterances

In this work, we present a novel means of browsing recorded audio conversations. The method we develop produces scalable summaries of the recognized speech, in which we can increase the amount of text continuously with the desired level of detail to best fill the available space. We present an interface in which a user can view an entire conversation in one screen, but can also quickly zoom in to see the full transcript; the corresponding audio can be easily played as well. The scaling is achieved via a combination of topic segmentation and informative phrase selection, where the threshold for informativeness decreases with increasing level of detail. Finally, we evaluate our method and interface against a baseline interface with a user study.

Due to speech recognition errors, repetitions are a frequent phenomenon in spoken dialog systems. In previous work (G. Zweig et al., 2008) we have proposed a joint decoding model that can leverage structural relationships between repeated utterances for improving recognition performance. In this paper we extend this work in two directions. First, we propose a direct, classification-based model for the same task. The new model can leverage features that were fundamentally hard to capture in the previous framework (e.g. spellings, false-starts, etc.) and leads to an additional performance improvement. Second, we show how both models can be used to perform a combined rescoring of two n-best lists that are part of a repetition pair.

We propose a technique for annotating data used to train a speech recognizer. The proposed scheme is based on labeling only a single frame for every word in the training set. We make use of the virtual evidence (VE) framework within a graphical model to take advantage of such data. We apply this approach to a large vocabulary speech recognition task, and show that our VE-based training scheme can improve over the performance of a system trained using sequence labeled data by 2.8% and 2.1% on the dev01 and eva101 sets respectively. Annotating data in the proposed scheme is not significantly slower than sequence labeling. We present timing results showing that training using the proposed approach is about 10 times faster than training using sequence labeled data while using only about 75% of the memory.

In this paper, we address the task of identifying the speakers by name in audio content. Identification of speakers by name helps to improve the readability of the transcript and also provides additional meta-data which can help in finding the audio content of interest. We present a conditional maximum entropy (maxent) framework for this problem which yields superior performance and lends itself well to incorporating different types of information. We take advantage of this property of maxent to explore new features for this task. We show that supplementing standard lexical triggers with information such as speaker gender and position of speaker name mentions afford us large gains in performance. At 95% precision, we increase the recall to 67% from the trigger baseline of 38%.

Speaker verification can be treated as a statistical hypothesis testing problem. The most commonly used approach is the likelihood ratio test (LRT), which can be shown to be optimal using the Neymann-Pearson lemma. However, in most practical situations the Neymann-Pearson lemma does not apply. In this paper, we present a more robust approach that makes use of a hybrid generative-discriminative framework for text-dependent speaker verification. Our algorithm makes use of a generative models to learn the characteristics of a speaker and then discriminative models to discriminate between a speaker and an impostor. One of the advantages of the proposed algorithm is that it does not require us to retrain the generative model. The proposed model, on an average, yields 36.41% relative improvement in EER over a LRT.

Voice-Rate is an automated dialog system which provides access to over one million ratings of products and businesses. By calling a toll-free number, consumers can access ratings for products, national businesses such as airlines, and local businesses such as restaurants. Voice-Rate also has a facility for recording and analyzing ratings that are given over the phone. The service has been primed with ratings taken from a variety of web sources, and we are augmenting these with user ratings. Voice-Rate can be accessed by dialing 1-877-456-DATA. 1

Multimedia content cannot be retrieved effectively unless metadata describing it is generated. However, metadata generation tends to be time-consuming and expensive, since it typically involves human beings going through the content and manually tagging it. The paper shows how automatic speech recognition (ASR) technology can be used to carry out metadata generation with significantly less expenditure of human effort. The paper describes two different approaches: voice tagging, whereby human beings tag the data but this process is speeded up by applying ASR to the tagging process; audio indexing, whereby much of the tagging process is automated by applying ASR to the content itself.

Speaker clustering is a key component in many speech processing applications. We focus on Broadcast News meta data annotation and speaker adaptation. In this setting, speaker clustering consists of identifying who spoke, and when they spoke in a long news broadcast. Speaker clustering is given a set of short audio segments. Ideally, it will discover how many people are speaking in the broadcast, and when they are speaking. The same problem can be transposed to a different domain. In this paper, we present two techniques that do not require a priori training. The speaker clustering is based on information collected solely on encountered test data. They aim at being portable across domains. The first method is based on a Bayesian information criterion (BIC), with single full-covariance Gaussians. It is fairly primitive but effective. The second method, called speaker triangulation, constructs a coordinate system based on conditional likelihoods of the audio segments. Clusters are located in this coordinate system. We are able to achieve state-of-the-art performance on NIST evaluations across different data sets.

This paper presents experiments of noise robust ASR on the Aurora4 database. The database is designed to test large vocabulary systems in presence of noise and channel distortions. A number of different model-based and signal-based noise robustness techniques have been tested. Results show that it is difficult to design a technique that is superior in every condition. Because of this we combined different techniques to improve results. Best results have been obtained when short time compensation / normalization methods are combined with long term environmental adaptation and robust acoustic models. The best average error rate obtained over the 52 conditions is 30.8%. This represents a 40 % relative improvement compared to the baseline results [1]. 1.

Setting linear constraints on HMM model space appears to be very effective for speaker adaptation. In doing so, we assume that model parameters are jointly Gaussian. While this approach has proven reasonably successful, we question it accuracy in the case of very high dimensionality parameter spaces. To address this problem, we employ a hierarchical piecewise linear model. Gross speaker variations are modeled with a linear eigenspace, subsuming the joint Gaussian model, and finer residues are modeled using another eigenspace chosen depending on the location of the first values. We perform experiments on Wall Street Journal (WSJ) dictation task, and we observe a cumulative 1.3% WER improvement (11 % relative) when using self-adaptation.

Cepstral mean normalization is the standard technique for channel robustness. Despite its good performance, the effectiveness of cepstral mean normalization (CMN) for short sentences is argued. CMN underlying hypothesis that the speech cepstral mean is constant is not valid for short processing windows. This implies the removal of some phonetic information. In this paper we show that the speech correlation structure may be used to estimate the communication channel and we propose an efficient algorithm to compute this estimate. We argue that the resulting channel estimate is more accurate because the underlying hypothesis is better verified than the original CMN hypothesis. Results for the Kai-Fu Lee phone recognition task on NTIMIT, with acoustic models trained on TIMIT (mismatch conditions), show that our method provides an 8% relative error rate reduction as compared to CMN.

The "eigenvoice" technique achieves rapid speaker adaptation by employing prior knowledge of speaker space obtained from reference speakers to place strong constraints on the initial model for each new speaker. It has previously been shown to yield very fast adaptation for a large-vocabulary system. In this paper, we describe a new way of applying the eigenvoice technique to context-dependent acoustic modeling, called the "eigencentroid plus delta trees" (EDT) model. Here, the context-dependent model is defined so that it consists of a speaker-dependent component with a small number of parameters linked to a speaker-independent component with far more parameters. The eigenvoice technique can then be applied to the speaker-dependent component alone to attain very fast adaptation of the entire context-dependent model (e.g., 10% relative reduction in error rate after 3 sentences). EDT requires only a small number of parameters to represent speaker space and works even if only a small amount of data is available per reference speaker.

We present an optimized implementation of the Viterbi algorithm suitable for small to large vocabulary, and isolated or continuous speech recognition. The Viterbi algorithm is certainly the most popular dynamic programming algorithm used in speech recognition. In this paper we propose a new algorithm that outperforms the Viterbi algorithm in term of complexity and of memory requirements. It is based on the assumption of strictly left to right models and explores the lexical tree in an optimal way, such that book-keeping computation is minimized. The tree is encoded such that children of a node are placed contiguously and in increasing order of memory heap so that the proposed algorithm also optimizes cache usage. Even though the algorithm is asymptotically two times faster that the conventional Viterbi algorithm, in our experiments

Gaussian Mixture Models (GMMs) have been successfully applied to the tasks of speaker ID and verification when a large amount of enrolment data is available to characterize client speakers ([1],[10],[11]). However, there are many applications where it is unreasonable to expect clients to spend this much time training the system. Thus, we have been exploring the performance of various methods when only a sparse amount of enrolment data is available. Under such conditions, the performance of GMMs deteriorates drastically. A possible solution is the “eigenvoice” approach, in which client and test speaker models are confined to a low-dimensional linear subspace obtained previously from a different set of training data. One advantage of the approach is that it does away with the need for impostor models for speaker verification. After giving a detailed description of the eigenvoice approach, the paper compares the performance of conventional GMMs on speaker ID and verification with that of GMMs obtained by means of the eigenvoice approach. Experimental results are presented to show that conventional GMMs perform better if there are abundant enrolment data, while eigenvoice GMMs perform better if enrolment data are sparse. The paper also gives experimental results for the case where the eigenspace is trained on one database (TIMIT), but client enrolment and testing involve another (YOHO). For this case, we show that performance improves if an environment adaptation technique is applied to the eigenspace. Finally, we discuss priorities for future work. training data client data train eigenspace estimate clients’ coordinates SPEAKER ID/VERIFICATION 1. use distance in eigenspace-> ‘‘eigendistance decoding’’ OR test data 2. use P(test data|client GMMs)-> ‘‘eigenGMM decoding’’ optional MLES re-estimation optional MLLR adaptation

This paper describes a new model-based speaker adaptation algorithm called the eigenvoice approach. The approach constrains the adapted model to be a linear combination of a small number of basis vectors obtained offline from a set of reference speakers, and thus greatly reduces the number of free parameters to be estimated from adaptation data. These "eigenvoice" basis vectors are orthogonal to each other and guaranteed to represent the most important components of variation between the reference speakers. Experimental results for a small-vocabulary task (letter recognition) given in the paper show that the approach yields major improvements in performance for tiny amounts of adaptation data. For instance, we obtained 16% relative improvement in error rate with one letter of supervised adaptation data, and 26% relative improvement with four letters of supervised adaptation data. After a comparison of the eigenvoice approach with other speaker adaptation algorithms, the paper concludes with a discussion of future work.

Previously, we presented a radically new class of fast adaptation techniques for speech recognition, based on prior knowledge of speaker variation. To obtain this prior knowledge, one applies a dimensionality reduction technique to T vectors of dimension D derived from T speaker-dependent (SD) models. This offline step yields T basis vectors, the eigenvoices. We constrain the model for new speaker S to be located in the space spanned by the first K eigenvoices. Speaker adaptation involves estimating K eigenvoice coefficients for the new speaker; typically, K is very small compared to original dimension D. Here, we review how to find the eigenvoices, give a maximum-likelihood estimator for the new speaker's eigenvoice coefficients, and summarize mean adaptation experiments carried out on the Isolet database. We present new results which assess the impact on performance of changes in training of the SD models. Finally, we interpret the first few eigenvoices obtained.

A new set of techniques exploiting N-best hypotheses in supervised and unsupervised adaptation are presented. These techniques combine statistics extracted from the N-best hypotheses with a weight derived from a likelihood ratio confidence measure. In the case of supervised adaptation the knowledge of the correct string is used to perform N-best based corrective adaptation. Experiments run for continuous letter recognition recorded in a car environment show that weighting N-best sequences by a likelihood ratio confidence measure provides only marginal improvement as compared to 1-best unsupervised adaptation and N-best unsupervised adaptation with equal weighting. However, an N-best based supervised corrective adaptation method weighting correct letters positively and incorrect letters negatively, resulted in a 13% decrease of the error rate as compared with supervised adaptation. The largest improvement was obtained for non-native speakers.

There are hidden analogies between two dissimilar research areas: face recognition and speech recognition. The standard representations for faces and voices misleadingly suggest that they have a high number of degrees of freedom. However, human faces have two eyes, a nose, and a mouth in predictable locations; such constraints ensure that possible images of faces occupy a tiny portion of the space of possible 2D images. Similarly, physical and cultural constraints on acoustic realizations of words uttered by a particular speaker imply that the true number of degrees of freedom for speaker-dependent hidden Markov models (HMMs) is quite small. Face recognition researchers have adopted representations that make explicit the underlying low dimensionality of the task, greatly improving the performance of their systems while reducing computational costs. We argue that speech researchers should use similar techniques to represent variation between speakers, and discuss applications to speaker adaptation, speaker identification and speaker verification.