Abstract Despite being the best known objective for learning speech representations, the HuBERT objective has not been further developed and improved. We argue that it is the lack of an underlying principle that stalls the development, and, in this paper, we show that predictive coding under a variational view is the principle behind the HuBERT objective. Due to its generality, our formulation provides opportunities to improve parameterization and optimization, and we show two simple modifications that bring immediate improvements to the HuBERT objective. In addition, the predictive coding formulation has tight connections to various other objectives, such as APC, CPC, wav2vec, and BEST-RQ. Empirically, the improvement in pre-training brings significant improvements to four downstream tasks: phone classification, f0 tracking, speaker recognition, and automatic speech recognition, highlighting the importance of the predictive coding interpretation.

Though producing an appropriate prosodic realisation of text is a one-to-many problem, modern speech generation often focuses on identifying the “best” or “most likely” output, overlooking acceptable variation across realisations. How listeners perceive such variation–and whether models capture it–is unaccounted for in current evaluation paradigms. In this study, we present exploratory analyses of whether self-supervised models encode acceptable prosodic variation. Using a new dataset of relative acceptability ratings across carefully controlled, high-quality synthetic utterances, we show that SSL representations contain information predictive of such judgments. By introducing a novel method for deriving probability-based uncertainty from autoregressive speech models, we examine whether this information is available in an unsupervised setting, highlighting the complexity of prosodic perception and the value of more human-centric evaluation paradigms.

Large language models (LLMs) have shown significant potential for robotics applications, particularly task planning, by harnessing their language comprehension and text generation capabilities. However, in applications such as household robotics, a critical gap remains in the personalization of these models to household preferences. For example, an LLM planner may find it challenging to perform tasks that require personalization, such as deciding where to place mugs in a kitchen based on specific household preferences. We introduce LLM-Personalize, a novel framework designed to personalize LLM planners for household robotics. LLM-Personalize uses an LLM planner to perform iterative planning in multi-room, partially-observable household environments, utilizing a scene graph built dynamically from local observations. To personalize the LLM planner towards user preferences, our optimization pipeline integrates imitation learning and reinforced Self-Training. We evaluate LLM-Personalize on Housekeep, a challenging simulated real-world 3D benchmark for household rearrangements, demonstrating a more than 30 percent increase in success rate over existing LLM planners, showcasing significantly improved alignment with human preferences.

Some End-to-End (E2E) Automatic Speech Recognition (ASR) models, such as Attention-based Encoder-Decoder (AED) and Recurrent Neural Network Transducer (RNN-T) are known to have components that effectively act as internal language models (ILM), implicitly modelling the prior probability of the output sequence. However, the existence of an ILM in pure Connectionist Temporal Classification (CTC) ASR systems remains debated. In this paper, we investigate the existence and strength of an ILM in CTC systems. Since CTC posterior probabilities cannot be analytically factorised, we propose a novel empirical method to probe the ILM. After validating our method on a hybrid DNN model with various external language models, we apply it to CTC models trained under different conditions, examining the effects of training data, modelling units, and training or pre-training methods. Our results show no strong evidence of an ILM in CTC-based ASR systems, even with the largest training dataset in our experiments. However, we make the surprising finding that when a CTC encoder is jointly trained with an AED loss, an ILM emerges, even when only the CTC component is used in decoding.

The lack of labeled data is a common challenge in speech classification tasks, particularly those requiring extensive subjective assessment, such as cognitive state classification. In this work, we propose a Semi-Supervised Learning (SSL) framework, introducing a novel multi-view pseudo-labeling method that leverages both acoustic and linguistic characteristics to select the most confident data for training the classification model. Acoustically, unlabeled data are compared to labeled data using the Fréchet audio distance, calculated from embeddings generated by multiple audio encoders. Linguistically, large language models are prompted to revise automatic speech recognition transcriptions and predict labels based on our proposed task-specific knowledge. High-confidence data are identified when pseudo-labels from both sources align, while mismatches are treated as low-confidence data. A bimodal classifier is then trained to iteratively label the low-confidence data until a predefined criterion is met. We evaluate our SSL framework on emotion recognition and dementia detection tasks. Experimental results demonstrate that our method achieves competitive performance compared to fully supervised learning using only 30% of the labeled data and significantly outperforms two selected baselines.

Annotating and recognizing speech emotion using prompt engineering has recently emerged with the advancement of Large Language Models (LLMs), yet its efficacy and reliability remain questionable. In this paper, we conduct a systematic study on this topic, beginning with the proposal of novel prompts that incorporate emotion-specific knowledge from acoustics, linguistics, and psychology. Subsequently, we examine the effectiveness of LLM-based prompting on Automatic Speech Recognition (ASR) transcription, contrasting it with ground-truth transcription. Furthermore, we propose a Revise-Reason-Recognize prompting pipeline for robust LLM-based emotion recognition from spoken language with ASR errors. Additionally, experiments on context-aware learning, in-context learning, and instruction tuning are performed to examine the usefulness of LLM training schemes in this direction. Finally, we investigate the sensitivity of LLMs to minor prompt variations. Experimental results demonstrate the efficacy of the emotion-specific prompts, ASR error correction, and LLM training schemes for LLM-based emotion recognition. Our study aims to refine the use of LLMs in emotion recognition and related domains.

We report on community-centered, collaborative research that weaves together HCI, natural language processing, linguistic, and design insights to develop spoken language technologies for unwritten languages. Across three visits to a Banjara farming community in India, we use participatory, technical, and creative methods to engage community members, collect spoken language photo annotations, and develop an information retrieval (IR) system. Drawing on orality theory, we interrogate assumptions and biases of current speech interfaces and create a simple application that leverages our IR system to match fluidly spoken queries with recorded annotations and surface corresponding photos. In-situ evaluations show how our novel approach returns reliable results and inspired the co-creation of media retrieval use-cases that are more appropriate in oral contexts. The very low (< 4h) spoken data requirements makes our approach adaptable to other contexts where languages are unwritten or have no digital language resources available.

Self-supervised learning methods (SSL) have seen wide popularity for speech representation learning. Early methods, such as wav2vec, were causal, whilst more recent approaches, notably wav2vec 2.0 and data2vec, have employed masking strategies together with a Transformer architecture. Many SSL methods use contrastive learning; however, non-contrastive methods, while susceptible to representational collapse, have recently seen success in other fields and have been successfully applied to speech in data2vec. This work returns to the study of causal models, comparing them with equivalent non-causal variants, motivated by our observation that non-contrastive SSL models have never been investigated in this setting. To this end, we propose a novel approach, Bootstrapped Predictive Coding (BPC), a causal non-contrastive SSL model. We find that causal SSL models outperform their non-causal counterparts in both contrastive and non-contrastive training setups, and that representations obtained with BPC give overall best performance when evaluated on phone frame classification.

Explainable AI (XAI) techniques have been widely used to help explain and understand the output of deep learning models in fields such as image classification and Natural Language Processing. Interest in using XAI techniques to explain deep learning-based Automatic Speech Recognition (ASR) is emerging. But there is not enough evidence on whether these explanations can be trusted. To address this, we adapt a state-of-the-art XAI technique from the image classification domain, Local Interpretable Model-Agnostic Explanations (LIME), to a model trained for a TIMIT-based phoneme recognition task. This simple task provides a controlled setting for evaluation while also providing expert annotated ground truth to assess the quality of explanations. We find a variant of LIME based on time partitioned audio segments, that we propose in this paper, produces the most reliable explanations, containing the ground truth 96% of the time in its top three audio segments.

With handling code-switching becoming an increasingly important topic in speech technology, driven by the expansion of low-resource and multilingual methodologies, it is vital that we recognize the diversity of code-switching as a phenomenon. We propose a framework that leverages linguistic findings as makeshift ground-truths to assess the quality and sufficiency of existing metrics designed to capture data-sets' differing code-switching styles. We also introduce a new metric, T-index, which leverages machine translation systems to capture properties of code-switched words in relation to the participating language pair. Through analysis of diverse Hindi-English and Mandarin-English datasets, we systematically explore how well these metrics align with linguistic intuition regarding code-switching richness levels in conversational versus technical domains.

Text data is commonly utilized as a primary input to enhance Speech Emotion Recognition (SER) performance and reliability. However, the reliance on human-transcribed text in most studies impedes the development of practical SER systems, creating a gap between in-lab research and real-world scenarios where Automatic Speech Recognition (ASR) serves as the text source. Hence, this study benchmarks SER performance using ASR transcripts with varying Word Error Rates (WERs) from eleven models on three well-known corpora: IEMOCAP, CMU-MOSI, and MSP-Podcast. Our evaluation includes both text-only and bimodal SER with six fusion techniques, aiming for a comprehensive analysis that uncovers novel findings and challenges faced by current SER research. Additionally, we propose a unified ASR error-robust framework integrating ASR error correction and modality-gated fusion, achieving lower WER and higher SER results compared to the best-performing ASR transcript. These findings provide insights into SER with ASR assistance, especially for real-world applications.

ASR remains unsatisfactory in scenarios where the speaking style diverges from that used to train ASR systems, resulting in erroneous transcripts. To address this, ASR Error Correction (AEC), a post-ASR processing approach, is required. In this work, we tackle an understudied issue: the Low-Resource Out-of-Domain (LROOD) problem, by investigating crossmodal AEC on very limited downstream data with 1-best hypothesis transcription. We explore pretraining and fine-tuning strategies and uncover an ASR domain discrepancy phenomenon, shedding light on appropriate training schemes for LROOD data. Moreover, we propose the incorporation of discrete speech units to align with and enhance the word embeddings for improving AEC quality. Results from multiple corpora and several evaluation metrics demonstrate the feasibility and efficacy of our proposed AEC approach on LROOD data as well as its generalizability and superiority on large-scale data. Finally, a study on speech emotion recognition confirms that our model produces ASR error-robust transcripts suitable for downstream applications.

Many recently published Text-to-Speech (TTS) systems produce audio close to real speech. However, TTS evaluation needs to be revisited to make sense of the results obtained with the new architectures, approaches and datasets. We propose evaluating the quality of synthetic speech as a combination of multiple factors such as prosody, speaker identity, and intelligibility. Our approach assesses how well synthetic speech mirrors real speech by obtaining correlates of each factor and measuring their distance from both real speech datasets and noise datasets. We benchmark 35 TTS systems developed between 2008 and 2024 and show that our score computed as an unweighted average of factors strongly correlates with the human evaluations from each time period.

Self-supervised learning (SSL) is used in deep learning to train on large datasets without the need for expensive labelling of the data. Recently, large Automatic Speech Recognition (ASR) models such as XLS-R have utilised SSL to train on over one hundred different languages simultaneously. However, deeper investigation shows that the bulk of the training data for XLS-R comes from a small number of languages. Biases learned through SSL have been shown to exist in multiple domains, but language bias in multilingual SSL ASR has not been thoroughly examined. In this paper, we utilise the Lottery Ticket Hypothesis (LTH) to identify language-specific subnetworks within XLS-R and test the performance of these subnetworks on a variety of different languages. We are able to show that when fine-tuning, XLS-R bypasses traditional linguistic knowledge and builds only on weights learned from the languages with the largest data contribution to the pretraining data.

Given recent advances in generative AI technology, a key question is how large language models (LLMs) can enhance acoustic modeling tasks using text decoding results from a frozen, pretrained automatic speech recognition (ASR) model. To explore new capabilities in language modeling for speech processing, we introduce the generative speech transcription error correction (GenSEC) challenge. This challenge comprises three post-ASR language modeling tasks: (i) post-ASR transcription correction, (ii) speaker tagging, and (iii) emotion recognition. These tasks aim to emulate future LLM-based agents handling voice-based interfaces while remaining accessible to a broad audience by utilizing open pretrained language models or agent-based APIs. We also discuss insights from baseline evaluations, as well as lessons learned for designing future evaluations.

Automatic Speech Recognition (ASR) systems often face challenges in alignment quality, particularly with the Connectionist Temporal Classification (CTC) approach, which frequently results in a high number of blank frames, known as the “peaky” issue. In this study, we explore the impact of modifying ASR model topologies on alignment quality without compromising Word Error Rate (WER) performance. Our findings demonstrate that introducing additional states to the CTC topology significantly improves alignment quality and mitigates the peaky issue. Conversely, increasing the minimum traversal frame can degrade alignment quality in our specific settings. These insights emphasise the critical importance of topology design in balancing alignment accuracy and recognition performance in ASR systems.

Traditionally, teaching a human and a Machine Learning (ML) model is quite different, but organized and structured learning has the ability to enable faster and better understanding of the underlying concepts. For example, when humans learn to speak, they first learn how to utter basic phones and then slowly move towards more complex structures such as words and sentences. Motivated by this observation, researchers have started to adapt this approach for training ML models. Since the main concept, the gradual increase in difficulty, resembles the notion of the curriculum in education, the methodology became known as Curriculum Learning (CL). In this work, we design and test new CL approaches to train Automatic Speech Recognition systems, specifically focusing on the so-called end-to-end models. These models consist of a single, large-scale neural network that performs the recognition task, in contrast to the traditional way of having several specialized components focusing on different subtasks (e.g., acoustic and language modeling). We demonstrate that end-to-end models can achieve better performances if they are provided with an organized training set consisting of examples that exhibit an increasing level of difficulty. To impose structure on the training set and to define the notion of an easy example, we explored multiple solutions that use either external, static scoring methods or incorporate feedback from the model itself. In addition, we examined the effect of pacing functions that control how much data is presented to the network during each training epoch. Our proposed curriculum learning strategies were tested on the task of speech recognition on two data sets, one containing spontaneous Finnish speech where volunteers were asked to speak about a given topic, and one containing planned English speech. Empirical results showed that a good curriculum strategy can yield performance improvements and speed-up convergence. After a given number of epochs, our best strategy achieved a 5.6% and 3.4% decrease in terms of test set word error rate for the Finnish and English data sets, respectively.

In this paper we develop approaches to automatic speech recognition (ASR) development that suit the needs and functions of under-heard language speakers. Our novel contribution to HCI is to show how community-engagement can surface key technical and social issues and opportunities for more effective speech-based systems. We introduce a bespoke toolkit of technologies and showcase how we utilised the toolkit to engage communities of under-heard language speakers; and, through that engagement process, situate key aspects of ASR development in community contexts. The toolkit consists of (1) an information appliance to facilitate spoken-data collection on topics of community interest, (2) a mobile app to create crowdsourced transcripts of collected data, and (3) demonstrator systems to showcase ASR capabilities and to feed back research results to community members. Drawing on the sensibilities we cultivated through this research, we present a series of challenges to the orthodoxy of state-of-the-art approaches to ASR development.

There has been a surge of interest regarding the alignment of large-scale language models with human language comprehension behaviour.The majority of this research investigates comprehension behaviours from reading isolated, written sentences.We propose studying the perception of dialogue, focusing on an intrinsic form of language use: spoken conversations.Using the task of predicting upcoming dialogue turns, we ask whether turn plausibility scores produced by state-of-the-art language models correlate with human judgements.We find a strong correlation for some but not all models: masked language models produce stronger correlations than autoregressive models.In doing so, we quantify human performance on the response selection task for open-domain spoken conversation.To the best of our knowledge, this is the first such quantification.We find that response selection performance can be used as a coarse proxy for the strength of correlation with human judgements, however humans and models make different response selection mistakes.The model which produces the strongest correlation also outperforms human response selection performance.Through ablation studies, we show that pre-trained language models provide a useful basis for turn representations; however, finegrained contextualisation, inclusion of dialogue structure information, and fine-tuning towards response selection all boost response selection accuracy by over 30 absolute points.

As a sub-branch of affective computing, impression recognition, e.g., perception of speaker characteristics such as warmth or competence, is potentially a critical part of both human-human conversations and spoken dialogue systems. Most research has studied impressions only from the behaviors expressed by the speaker or the response from the listener, yet ignored their latent connection. In this paper, we perform impression recognition using a proposed listener adaptive cross-domain architecture, which consists of a listener adaptation function to model the causality between speaker and listener behaviors and a cross-domain fusion function to strengthen their connection. The experimental evaluation on the dyadic IMPRESSION dataset verified the efficacy of our method, producing concordance correlation coefficients of 78.8% and 77.5% in the competence and warmth dimensions, outperforming previous studies. The proposed method is expected to be generalized to similar dyadic interaction scenarios in affective computing.

English is the most widely spoken language in the world, used daily by millions of people as a first or second language in many different contexts. As a result, there are many varieties of English. Although the great many advances in English automatic speech recognition (ASR) over the past decades, results are usually reported based on test datasets which fail to represent the diversity of English as spoken today around the globe. We present the first release of The Edinburgh International Accents of English Corpus (EdAcc). This dataset attempts to better represent the wide diversity of English, encompassing almost 40 hours of dyadic video call conversations between friends. Unlike other datasets, EdAcc includes a wide range of first and second-language varieties of English and a linguistic background profile of each speaker. Results on latest public, and commercial models show that EdAcc highlights shortcomings of current English ASR models. The best performing model, trained on 680 thousand hours of transcribed data, obtains an average of 19.7% word error rate (WER) – in contrast to the 2.7% WER obtained when evaluated on US English clean read speech. Across all models, we observe a drop in performance on Indian, Jamaican, and Nigerian English speakers. Recordings, linguistic backgrounds, data statement, and evaluation scripts are released on our website under CC-BY-SA1license.2We hope that this work will encourage future research on a wider range of English varieties to create more accessible speech technologies.

We propose a new method for human speech intelligibility evaluation based on keyword spotting. In this method, participants play a stimulus and select the word they hear from a close set of alternatives. To find which sentence to use, the target word, and alternatives we mine a large set of stimuli using a phonetic dictionary and a language model. Unlike other tests, our method does not rely on specially designed sentences and can be used to evaluate in-the-wild material such as TED talks. We focus on audio-visual (AV) speech enhancement (SE) evaluation as a study case. We compared our method to a transcription task and observed that the two produce highly correlated results, albeit our task requiring substantially less participation time. We then adopted it on a large-scale evaluation of AVSE systems. Results show that keyword spotting is a suitable and efficient alternative to assess intelligibility from AV stimuli.

We address quality assessment for neural network based ASR by providing explanations that help increase our understanding of the system and ultimately help build trust in the system. Compared to simple classification labels, explaining transcriptions is more challenging as judging their correctness is not straightforward and transcriptions as a variable-length sequence is not handled by existing interpretable machine learning models.We provide an explanation for an ASR transcription as a subset of audio frames that is both a minimal and sufficient cause of the transcription. To do this, we adapt existing explainable AI (XAI) techniques from image classification - (1) Statistical Fault Localisation(SFL) [1] and (2) Causal [2]. Additionally, we use an adapted version of Local Interpretable Model-Agnostic Explanations (LIME) [3] for ASR as a baseline in our experiments. We evaluate the quality of the explanations generated by the proposed techniques over three different ASR – Google API [4], the baseline model of Sphinx [5], Deepspeech [6] – and 100 audio samples from the Commonvoice dataset [7].

The linguistic notion of formality is one dimension of stylistic variation in human communication. A universal characteristic of language production, formality has surface-level realizations in written and spoken language. In this work, we explore ways of measuring the formality of such realizations in multilingual speech corpora across a wide range of domains. We compare measures of formality, contrasting textual and acoustic-prosodic metrics. We believe that a combination of these should correlate well with downstream applications. Our findings include: an indication that certain prosodic variables might play a stronger role than others; no correlation between prosodic and textual measures; limited evidence for anticipated inter-domain trends, but some evidence of consistency of measures between languages. We conclude that non-lexical indicators of formality in speech may be more subtle than our initial expectations, motivating further work on reliably encoding spoken formality.

Speech is a fundamental means of communication that can be seen to provide two channels for transmitting information: the lexical channel of which words are said, and the non-lexical channel of how they are spoken. Both channels shape listener expectations of upcoming communication; however, directly quantifying their relative effect on expectations is challenging. Previous attempts require spoken variations of lexically equivalent dialogue turns or conspicuous acoustic manipulations. This paper introduces a generalised paradigm to study the value of non-lexical information in dialogue across unconstrained lexical content. By quantifying the perceptual value of the non-lexical channel with both accuracy and entropy reduction, we show that non-lexical information produces a consistent effect on expectations of upcoming dialogue: even when it leads to poorer discriminative turn judgements than lexical content alone, it yields higher consensus among participants.

Holistic perception of affective attributes is an important human perceptual ability. However, this ability is far from being realized in current affective computing, as not all of the attributes are well studied and their interrelationships are poorly understood. In this work, we investigate the relationship between two affective attributes: personality and emotion, from a transfer learning perspective. Specifically, we transfer Transformer-based and wav2vec-based emotion recognition models to perceive personality from speech across corpora. Compared with previous studies, our results show that transferring emotion recognition is effective for personality perception. Moreoever, this allows for better use and exploration of small personality corpora. We also provide novel findings on the relationship between personality and emotion that will aid future research on holistic affect recognition.

In Speech Emotion Recognition (SER), textual data is often used alongside audio signals to address their inherent variability. However, the reliance on human annotated text in most research hinders the development of practical SER systems. To overcome this challenge, we investigate how Automatic Speech Recognition (ASR) performs on emotional speech by analyzing the ASR performance on emotion corpora and examining the distribution of word errors and confidence scores in ASR transcripts to gain insight into how emotion affects ASR. We utilize four ASR systems, namely Kaldi ASR, wav2vec, Conformer, and Whisper, and three corpora: IEMOCAP, MOSI, and MELD to ensure generalizability. Additionally, we conduct text-based SER on ASR transcripts with increasing word error rates to investigate how ASR affects SER. The objective of this study is to uncover the relationship and mutual impact of ASR and SER, in order to facilitate ASR adaptation to emotional speech and the use of SER in real world.

End-to-end (E2E) Automatic Speech Recognition (ASR) has gained popularity in recent years, with most research focusing on designing novel neural network architectures, speech rep resentations, and loss functions. However, the importance of topology in E2E ASR has been largely neglected. There are many aspects of topology to consider; in this paper, we focus on the relationship between topologies’ minimum traversal time and output frame rate, the number of distinct states for each output unit, and the flexibility of alignments admitted. We ex amine several different topologies on two datasets: WSJ and Librispeech. Our experiments reveal that different frame rates have varying optimal topologies and that the commonly used Connectionist Temporal Classification (CTC) topology is not always optimal. Our findings suggest that the choice of topol ogy is an important consideration in the design of E2E ASR systems.

Speech is a fundamental means of communication that can be seen to provide two channels for transmitting information: the lexical channel of \textit{which} words are said, and the non-lexical channel of \textit{how} they are spoken. Both channels shape listener expectations of upcoming communication; however, directly quantifying their relative effect on expectations is challenging. Previous attempts require spoken variations of lexically-equivalent dialogue turns or conspicuous acoustic manipulations. This paper introduces a generalised paradigm to study the value of non-lexical information in dialogue across unconstrained lexical content.By quantifying the perceptual value of the non-lexical channel with both accuracy and entropy reduction, we show that non-lexical information produces a consistent effect on expectations of upcoming dialogue: even when it leads to poorer discriminative turn judgements than lexical content alone, it yields higher consensus among participants.

End-to-end (E2E) Automatic Speech Recognition (ASR) has gained popularity in recent years, with most research focusing on designing novel neural network architectures, speech rep resentations, and loss functions. However, the importance of topology in E2E ASR has been largely neglected. There are many aspects of topology to consider; in this paper, we focus on the relationship between topologies’ minimum traversal time and output frame rate, the number of distinct states for each output unit, and the flexibility of alignments admitted. We ex amine several different topologies on two datasets: WSJ and Librispeech. Our experiments reveal that different frame rates have varying optimal topologies and that the commonly used Connectionist Temporal Classification (CTC) topology is not always optimal. Our findings suggest that the choice of topol ogy is an important consideration in the design of E2E ASR systems.

In this paper, we analyse the performance of the TIMIT-based phone recognition systems beyond the overall phone error rate (PER) metric. We consider three broad phonetic classes (BPCs): {affricate, diphthong, fricative, nasal, plosive, semi-vowel, vowel, silence}, {consonant, vowel, silence} and {voiced, unvoiced, silence} and, calculate the contribution of each phonetic class in terms of the substitution, deletion, insertion and PER. Furthermore, for each BPC we investigate the following: evolution of PER during training, effect of noise (NTIMIT), importance of different spectral subbands (1, 2, 4, and 8 kHz), usefulness of bidirectional vs unidirectional sequential modelling, transfer learning from WSJ and regularisation via monophones. In addition, we construct a confusion matrix for each BPC and analyse the confusions via dimensionality reduction to 2D at the input (acoustic features) and output (logits) levels of the acoustic model. We also compare the performance and confusion matrices of the BLSTM-based hybrid baseline system with those of the GMM-HMM based hybrid, Conformer and wav2vec 2.0 based end-to-end phone recognisers. Finally, the relationship of the unweighted and weighted PERs with the broad phonetic class priors is studied for both the hybrid and end-to-end systems.

In this paper, we investigate multi-stream acoustic modelling using the raw real and imaginary parts of the Fourier transform of speech signals. Using the raw magnitude spectrum, or features derived from it, as a proxy for the real and imaginary parts leads to irreversible information loss and suboptimal information fusion. We discuss and quantify the importance of such information in terms of speech quality and intelligibility. In the proposed framework, the real and imaginary parts are treated as two streams of information, pre-processed via separate convolutional networks, and then combined at an optimal level of abstraction, followed by further post-processing via recurrent and fully-connected layers. The optimal level of information fusion in various architectures, training dynamics in terms of cross-entropy loss, frame classification accuracy and WER as well as the shape and properties of the filters learned in the first convolutional layer of single- and multi-stream models are analysed. We investigated the effectiveness of the proposed systems in various tasks: TIMIT/NTIMIT (phone recognition), Aurora-4 (noise robustness), WSJ (read speech), AMI (meeting) and TORGO (dysarthric speech). Across all tasks we achieved competitive performance: in Aurora-4, down to 4.6% WER on average, in WSJ down to 4.6% and 6.2% WERs for Eval-92 and Eval-93, for Dev/Eval sets of the AMI-IHM down to 23.3%/23.8% WERs and in the AMI-SDM down to 43.7%/47.6% WERs have been achieved. In TORGO, for dysarthric and typical speech we achieved down to 31.7% and 10.2% WERs, respectively.

Automatic Speech Recognition (ASR) researchers are turning their attention towards supporting low-resource languages, such as isiXhosa or Marathi, with only limited training resources. We report and reflect on collaborative research across ASR & HCI to situate ASR-enabled technologies to suit the needs and functions of two communities of low-resource language speakers, on the outskirts of Cape Town, South Africa and in Mumbai, India. We build on longstanding community partnerships and draw on linguistics, media studies and HCI scholarship to guide our research. We demonstrate diverse design methods to: remotely engage participants; collect speech data to test ASR models; and ultimately field-test models with users. Reflecting on the research, we identify opportunities, challenges, and use-cases of ASR, in particular to support pervasive use of WhatsApp voice messaging. Finally, we uncover implications for collaborations across ASR & HCI that advance important discussions at CHI surrounding data, ethics, and AI.

This paper seeks to improve the performance of automatic speech recognition (ASR) systems operating on code-switched speech. Code-switching refers to the alternation of languages within a conversation, a phenomenon that is of increasing importance considering the rapid rise in the number of bilingual speakers in the world. It is particularly challenging for ASR owing to the relative scarcity of code-switching speech and text data, even when the individual languages are themselves well-resourced. This paper proposes to overcome this challenge by applying linguistic theories in order to generate more realistic code-switching text, necessary for language modelling in ASR. Working with English-Spanish code-switching, we find that Equivalence Constraint theory and part-of-speech labelling are particularly helpful for text generation, and bring 2% improvement to ASR performance.

Alongside acoustic information, linguistic features based on speech transcripts have been proven useful in Speech Emotion Recognition (SER). However, due to the scarcity of emotion labelled data and the difficulty of recognizing emotional speech, it is hard to obtain reliable linguistic features and models in this research area. In this paper, we propose to fuse Automatic Speech Recognition (ASR) outputs into the pipeline for joint training SER. The relationship between ASR and SER is understudied, and it is unclear what and how ASR features benefit SER. By examining various ASR outputs and fusion methods, our experiments show that in joint ASR-SER training, incorporating both ASR hidden and text output using a hierarchical co-attention fusion approach improves the SER performance the most. On the IEMOCAP corpus, our approach achieves 63.4% weighted accuracy, which is close to the baseline results achieved by combining ground-truth transcripts. In addition, we also present novel word error rate analysis on IEMOCAP and layer-difference analysis of the Wav2vec 2.0 model to better understand the relationship between ASR and SER.

End-to-end Automatic Speech Recognition (E2E ASR) significantly simplifies the training process of an ASR model. Connectionist Temporal Classification (CTC) is one of the most popular methods for E2E ASR training. Implicitly, CTC has a unique topology which is very useful for sequence modelling. However, we find that by changing to another topology, we can make it even more effective. In this paper, we propose a new CTC-like method, for E2E ASR training, by modifying the topology of original CTC, so that the well-known abuse of the blank label in CTC can be resolved theoretically. As we change the topology, a normalisation term is necessary, which makes the form of the final loss function similar to Maximum Mutual Information (MMI); we hence name our method MMI-CTC. In addition to maximising the posterior probability of the target sequence, the normalisation enables models to explicitly minimise the probability of competing hypothesis at the word sequence level. Our experimental results show that MMI-CTC is more efficient than CTC, and that the normalisation is essential for sequence training.

We present a method for cross-lingual training an ASR system using absolutely no transcribed training data from the target language, and with no phonetic knowledge of the language in question. Our approach uses a novel application of a decipherment algorithm, which operates given only unpaired speech and text data from the target language. We apply this decipherment to phone sequences generated by a universal phone recogniser trained on out-of-language speech corpora, which we follow with flat-start semi-supervised training to obtain an acoustic model for the new language. To the best of our knowledge, this is the first practical approach to zero-resource cross-lingual ASR which does not rely on any hand-crafted phonetic information. We carry out experiments on read speech from the GlobalPhone corpus, and show that it is possible to learn a decipherment model on just 20 minutes of data from the target language. When used to generate pseudo-labels for semi-supervised training, we obtain WERs that range from 32.5% to just 1.9% absolute worse than the equivalent fully supervised models trained on the same data.

Deep speaker embeddings have been shown to encode a wide variety of attributes relating to a speaker. The aim of this work is to separate out some of these attributes in the embedding space, disentangling these sources of speaker variation into subsets of the embedding dimensions. This is achieved modifying the training procedure of a typical speaker embedding network, which is typically only trained to classify speakers. This work instead adds pairs of attribute specific task heads to operate on complementary subsets of the speaker embedding dimensions. While specific dimensions are encouraged to encode an attribute, for example gender, the other dimensions are penalized for containing this information using an adversarial loss. We show that this method is effective in factorizing out multiple attributes in the embedding space, successfully disentangling gender, nationality and age. Using the disentangled representations, we investigate how much removing this information impacts speaker verification and diarization performance, showing that gender is a significant source of separation in the deep speaker embedding space, with nationality and age also contributing to a lesser degree.

Surprisal is used throughout computational psycholinguistics to model a range of language processing behaviour. There is growing evidence that language model (LM) estimates of surprisal correlate with human performance on a range of written language comprehension tasks. Although communicative interaction is arguably the primary form of language use, most studies of surprisal are based on monological, written data. Towards the goal of understanding perception in spontaneous, natural language, we present an exploratory investigation into whether the relationship between human comprehension behaviour and LM-estimated surprisal holds when applied to dialogue, considering both written dialogue, and the lexical component of spoken dialogue. We use a novel judgement task of dialogue utterance acceptability to ask two questions: “How well can people make predictions about written dialogue and transcripts of spoken dialogue?” and “Does surprisal correlate with these acceptability judgements?”. We demonstrate that people can make accurate predictions about upcoming dialogue and that their ability differs between spoken transcripts and written conversation. We investigate the relationship between global and local operationalisations of surprisal and human acceptability judgements, finding a combination of both to provide the most predictive power

Audio-visual speech enhancement is the task of improving the quality of a speech signal when video of the speaker is available. It opens-up the opportunity of improving speech intelligibility in adverse listening scenarios that are currently too challenging for audio-only speech enhancement models. The Audio-Visual Speech Enhancement (AVSE) challenge aims to set the first benchmark in this area. We provide participants with datasets and scripts to test their audio-visual speech enhancement models under a common framework for both training and evaluation. The data is derived from real-world videos, and comprises noisy mixes, in which audio from target speaker is mixed with either a competing speaker or a noise signal. The submitted systems are evaluated by conducting AV intelligibility tests involving human participants. We expect this challenge to be a platform for advancing the field of audio-visual speech-enhancement and to provide further insight about the scope and limitations of current AV speech enhancement approaches.

Self-supervised speech models have grown fast during the past few years and have proven feasible for use in various downstream tasks. Some recent work has started to look at the characteristics of these models, yet many concerns have not been fully addressed. In this work, we conduct a study on emotional corpora to explore a popular self-supervised model - wav2vec 2.0. Via a set of quantitative analysis, we mainly demonstrate that: 1) wav2vec 2.0 appears to discard paralinguistic information that is less useful for word recognition purposes; 2) for emotion recognition, representations from the middle layer alone perform as well as those derived from layer averaging, while the final layer results in the worst performance in some cases; 3) current self-supervised models may not be the optimal solution for downstream tasks that make use of non-lexical features. Our work provides novel findings that will aid future research in this area and theoretical basis for the use of existing models.

Acoustic models are susceptible to the difference in acoustic characteristics between the training distribution and test distributions. Accent variability is a challenging source of variability, and the variations within one accent often do not generalize to others. Consequently, end-to-end models that have only transcriptions as linguistic information need high amounts of data to learn how different accents realize their sounds. To aid with recognition of accented speech, we make use of an accent independent abstraction of phonemes, often called metaphonemes. We force our models to learn hidden representations that are correlated to metaphonemes using multi-task training. Our aim is to obtain a model that is more robust to accented speech and, can, at the same time, adapt faster to different accents through the learned structure. Our experiments on the Common Voice corpus show better generalization when making use of this additional linguistic information, with a word error rate reduction of up to 12.6% when compared to the baseline. Furthermore, the relative improvement when adapting an existing model by making use of the metaphonemes is higher than using Byte Pair Encodings alone.

David Wan, Chris Kedzie, Faisal Ladhak, Elsbeth Turcan, Petra Galuscakova, Elena Zotkina, Zhengping Jiang, Peter Bell, Kathleen McKeown. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.

Magnitude spectrum-based features are the most widely employed front-ends for acoustic modelling in automatic speech recognition (ASR) systems. In this paper, we investigate the possibility and efficacy of acoustic modelling using the raw short-time phase spectrum. In particular, we study the usefulness of the raw wrapped, unwrapped and minimum-phase phase spectra as well as the phase of the source and filter components for acoustic modelling. Furthermore, we explore the effectiveness of simultaneous deployment of the vocal tract and excitation components of the raw phase spectrum using multi-head CNNs and investigate multiple information fusion schemes. This paves the way for developing an effective phase-based multi-stream information processing systems for speech recognition. The performance, even for wrapped phase with a noise-like shape, is comparable to or better than the magnitude-based classic features, and up to 4.8% WER has been achieved in the WSJ (Eval-92) task.

Although the lower layers of a deep neural network learn features which are transferable across datasets, these layers are not transferable within the same dataset. That is, in general, freezing the trained feature extractor (the lower layers) and retraining the classifier (the upper layers) on the same dataset leads to worse performance. In this paper, for the first time, we show that the frozen classifier is transferable within the same dataset. We develop a novel top-down training method which can be viewed as an algorithm for searching for high-quality classifiers. We tested this method on automatic speech recognition (ASR) tasks and language modelling tasks. The proposed method consistently improves recurrent neural network ASR models on Wall Street Journal, self-attention ASR models on Switchboard, and AWD-LSTM language models on WikiText-2.

This paper describes the CSTR submission to the Multilingual and Code-Switching ASR Challenges at Interspeech 2021. For the multilingual track of the challenge, we trained a multilingual CNN-TDNN acoustic model for Gujarati, Hindi, Marathi, Odia, Tamil and Telugu and subsequently fine-tuned the model on monolingual training data. A language model built on a mixture of training and CommonCrawl data was used for decoding. We also demonstrate that crawled data from YouTube can be successfully used to improve the performance of the acoustic model with semi-supervised training. These models together with confidence based language identification achieve the average WER of 18.1%, a 41% relative improvement compared to the provided multilingual baseline model. For the code-switching track of the challenge we again train a multilingual model on Bengali and Hindi technical lectures and we employ a language model trained on CommonCrawl Bengali and Hindi data mixed with in-domain English data, using a novel transliteration method to generate pronunciations for the English terms. The final model improves by 18% and 34% relative compared to our multilingual baseline. Both our systems were among the top-ranked entries to the challenge.

Source-filter modelling is among the fundamental techniques in speech processing with a wide range of applications. In acoustic modelling, features such as MFCC and PLP which parametrise the filter component are widely employed. In this paper, we investigate the efficacy of building acoustic models from the raw filter and source components. The raw magnitude spectrum, as the primary information stream, is decomposed into the excitation and vocal tract information streams via cepstral liftering. Then, acoustic models are built via multi-head CNNs which, among others, allow for processing each individual stream via a sequence of bespoke transforms and fusing them at an optimal level of abstraction. We discuss the possible advantages of such information factorisation and recombination, investigate the dynamics of these models and explore the optimal fusion level. Furthermore, we illustrate the CNN’s learned filters and provide some interpretation for the captured patterns. The proposed approach with optimal fusion scheme results in up to 14% and 7% relative WER reduction in WSJ and Aurora-4 tasks.

Deep speaker embeddings have become the leading method for encoding speaker identity in speaker recognition tasks. The embedding space should ideally capture the variations between all possible speakers, encoding the multiple acoustic aspects that make up a speaker's identity, whilst being robust to non-speaker acoustic variation. Deep speaker embeddings are normally trained discriminatively, predicting speaker identity labels on the training data. We hypothesise that additionally predicting speaker-related auxiliary variables -- such as age and nationality -- may yield representations that are better able to generalise to unseen speakers. We propose a framework for making use of auxiliary label information, even when it is only available for speech corpora mismatched to the target application. On a test set of US Supreme Court recordings, we show that by leveraging two additional forms of speaker attribute information derived respectively from the matched training data, and VoxCeleb corpus, we improve the performance of our deep speaker embeddings for both verification and diarization tasks, achieving a relative improvement of 26.2% in DER and 6.7% in EER compared to baselines using speaker labels only. This improvement is obtained despite the auxiliary labels having been scraped from the web and being potentially noisy.

People convey information extremely effectively through spoken interaction using multiple channels of information transmission: the lexical channel of what is said, and the non-lexical channel of how it is said. We propose studying human perception of spoken communication as a means to better understand how information is encoded across these channels, focusing on the question 'What characteristics of communicative context affect listener's expectations of speech?'. To investigate this, we present a novel behavioural task testing whether listeners can discriminate between the true utterance in a dialogue and utterances sampled from other contexts with the same lexical content. We characterize how perception - and subsequent discriminative capability - is affected by different degrees of additional contextual information across both the lexical and non-lexical channel of speech. Results demonstrate that people can effectively discriminate between different prosodic realisations, that non-lexical context is informative, and that this channel provides more salient information than the lexical channel, highlighting the importance of the non-lexical channel in spoken interaction.

The use of semi-supervised training (SST) has become an increasingly popular way of increasing the performance of ASR acoustic models without the need for further transcribed speech data. However, the performance of the technique can be very sensitive to the quality of the initial ASR system. This paper undertakes a comprehensive study of the improvements gained with respect to variation in the initial systems, the quantity of untranscribed data used, and the learning schedules. We postulate that the reason SST can be effective even when the initial model is poor is because it enables utterance-level information to be propagated to the frame level, and hence hypothesise that the quality of the language model plays a much larger role than the quality of the acoustic model. In experiments on Tagalog data from the IARPA MATERIAL programme, we find that indeed this is the case, and show that with an appropriately chosen recipe it is possible to achieve over 50% relative WER reductions from SST, even when the WER of the initial system is more than 80%.

Recently, Transformer based models have shown competitive automatic speech recognition (ASR) performance. One key factor in the success of these models is the multi-head attention mechanism. However, for trained models, we have previously observed that many attention matrices are close to diagonal, indicating the redundancy of the corresponding attention heads. We have also found that some architectures with reduced numbers of attention heads have better performance. Since the search for the best structure is time prohibitive, we propose to randomly remove attention heads during training and keep all attention heads at test time, thus the final model is an ensemble of models with different architectures. The proposed method also forces each head independently learn the most useful patterns. We apply the proposed method to train Transformer based and Convolution-augmented Transformer (Conformer) based ASR models. Our method gives consistent performance gains over strong baselines on the Wall Street Journal, AISHELL, Switchboard and AMI datasets. To the best of our knowledge, we have achieved state-of-the-art end-to-end Transformer based model performance on Switchboard and AMI.

Self-attention models such as Transformers, which can capture temporal relationships without being limited by the distance between events, have given competitive speech recognition results. However, we note the range of the learned context increases from the lower to upper self-attention layers, whilst acoustic events often happen within short time spans in a left-to-right order. This leads to a question: for speech recognition, is a global view of the entire sequence useful for the upper self-attention encoder layers in Transformers? To investigate this, we train models with lower self-attention/upper feed-forward layers encoders on Wall Street Journal and Switchboard. Compared to baseline Transformers, no performance drop but minor gains are observed. We further developed a novel metric of the diagonality of attention matrices and found the learned diagonality indeed increases from the lower to upper encoder self-attention layers. We conclude the global view is unnecessary in training upper encoder layers.

We propose a method for zero-resource domain adaptation of DNN acoustic models, for use in low-resource situations where the only in-language training data available may be poorly matched to the intended target domain. Our method uses a multi-lingual model in which several DNN layers are shared between languages. This architecture enables domain adaptation transforms learned for one well-resourced language to be applied to an entirely different low- resource language. First, to develop the technique we use English as a well-resourced language and take Spanish to mimic a low-resource language. Experiments in domain adaptation between the conversational telephone speech (CTS) domain and broadcast news (BN) domain demonstrate a 29% relative WER improvement on Spanish BN test data by using only English adaptation data. Second, we demonstrate the effectiveness of the method for low-resource languages with a poor match to the well-resourced language. Even in this scenario, the proposed method achieves relative WER improvements of 18-27% by using solely English data for domain adaptation. Compared to other related approaches based on multi-task and multi-condition training, the proposed method is able to better exploit well-resource language data for improved acoustic modelling of the low-resource target domain.

Previous work has encouraged domain-invariance in deep speaker embedding by adversarially classifying the dataset or labelled environment to which the generated features belong. We propose a training strategy which aims to produce features that are invariant at the granularity of the recording or channel, a finer grained objective than dataset- or environment-invariance. By training an adversary to predict whether pairs of same-speaker embeddings belong to the same recording in a Siamese fashion, learned features are discouraged from utilizing channel information that may be speaker discriminative during training. Experiments for verification on VoxCeleb and diarization and verification on CALLHOME show promising improvements over a strong baseline in addition to outperforming a dataset-adversarial model. The VoxCeleb model in particular performs well, achieving a 4% relative improvement in EER over a Kaldi baseline, while using a similar architecture and less training data.

We propose a multi-scale octave convolution layer to learn robust speech representations efficiently. Octave convolutions were introduced by Chen et al [1] in the computer vision field to reduce the spatial redundancy of the feature maps by decomposing the output of a convolutional layer into feature maps at two different spatial resolutions, one octave apart. This approach improved the efficiency as well as the accuracy of the CNN models. The accuracy gain was attributed to the enlargement of the receptive field in the original input space. We argue that octave convolutions likewise improve the robustness of learned representations due to the use of average pooling in the lower resolution group, acting as a low-pass filter. We test this hypothesis by evaluating on two noisy speech corpora - Aurora-4 and AMI. We extend the octave convolution concept to multiple resolution groups and multiple octaves. To evaluate the robustness of the inferred representations, we report the similarity between clean and noisy encodings using an affine projection loss as a proxy robustness measure. The results show that proposed method reduces the WER by up to 6.6% relative for Aurora-4 and 3.6% for AMI, while improving the computational efficiency of the CNN acoustic models.

The central auditory pathway exploits the auditory signals and visual information sent by both ears and eyes to segregate speech from multiple competing noise sources and help disambiguate phonological ambiguity. In this study, inspired from this unique human ability, we present a deep neural network (DNN) that ingest the binaural sounds received at the two ears as well as the visual frames to selectively suppress the competing noise sources individually at both ears. The model exploits the noisy binaural cues and noise robust visual cues to improve speech intelligibility. The comparative simulation results in terms of objective metrics such as PESQ, STOI, SI-SDR and DBSTOI demonstrate significant performance improvement of the proposed audio-visual (AV) DNN as compared to the audio-only (A-only) variant of the proposed model. Finally, subjective listening tests with the real noisy AV ASPIRE corpus shows the superiority of the proposed AV DNN as compared to state-of-the-art approaches.

Deep scattering spectrum consists of a cascade of wavelet transforms and modulus non-linearity. It generates features of different orders, with the first order coefficients approximately equal to the Mel-frequency cepstrum, and higher order coefficients recovering information lost at lower levels. We investigate the effect of including the information recovered by higher order coefficients on the robustness of speech recognition. To that end, we also propose a modification to the original scattering transform tailored for noisy speech. In particular, instead of the modulus non-linearity we opt to work with power coefficients and, therefore, use the squared modulus non-linearity. We quantify the robustness of scattering features using the word error rates of acoustic models trained on clean speech and evaluated using sets of utterances corrupted with different noise types. Our empirical results show that the second order scattering power spectrum coefficients capture invariants relevant for noise robustness and that this additional information improves generalization to unseen noise conditions (almost 20% relative error reduction on aurora 4). This finding can have important consequences on speech recognition systems that typically discard the second order information and keep only the first order features (known for emulating mfcc and fbank values) when representing speech.

We investigate the robustness and training dynamics of raw waveform acoustic models for automatic speech recognition (ASR). It is known that the first layer of such models learn a set of filters, performing a form of time-frequency analysis. This layer is liable to be under-trained owing to gradient vanishing, which can negatively affect the network performance. Through a set of experiments on TIMIT, Aurora-4 and WSJ datasets, we investigate the training dynamics of the first layer by measuring the evolution of its average frequency response over different epochs. We demonstrate that the network efficiently learns an optimal set of filters with a high spectral resolution and the dynamics of the first layer highly correlates with the dynamics of the cross entropy (CE) loss and word error rate (WER). In addition, we study the robustness of raw waveform models in both matched and mismatched conditions. The accuracy of these models is found to be comparable to, or better than, their MFCC-based counterparts in matched conditions and notably improved by using a better alignment. The role of raw waveform normalisation was also examined and up to 4.3% absolute WER reduction in mismatched conditions was achieved.

In this paper we investigate the usefulness of the sign spectrum and its combination with the raw magnitude spectrum in acoustic modelling for automatic speech recognition (ASR). The sign spectrum is a sequence of ±1s, capturing one bit of the phase spectrum. It encodes information overlooked by the magnitude spectrum enabling unique signal characterisation and reconstruction. In particular, we demonstrate it carries information related to the temporal structure of the signal as well as the speech’s source component. Furthermore, we investigate the usefulness of combining it with the raw magnitude spectrum via multi-head CNNs at different fusion levels for ASR. While information-wise these two streams of information are together equivalent to the raw waveform signal the overall performance is noticeably higher than raw waveform and classic features such as MFCC and filterbank. This has been observed and verified in TIMIT, NTIMT, Aurora-4 and WSJ tasks and up to 14.5% relative WER reduction has been achieved.

Due to limited computational resources, acoustic models of early automatic speech recognition ( asr) systems were built in low-dimensional feature spaces that incur considerable information loss at the outset of the process. Several comparative studies of automatic and human speech recognition suggest that this information loss can adversely affect the robustness of asr systems. To mitigate that and allow for learning of robust models, we propose a deep 2 d convolutional network in the waveform domain. The first layer of the network decomposes waveforms into frequency sub-bands, thereby representing them in a structured high-dimensional space. This is achieved by means of a parametric convolutional block defined via cosine modulations of compactly supported windows. The next layer embeds the waveform in an even higher-dimensional space of high-resolution spectro-temporal patterns, implemented via a 2 d convolutional block. This is followed by a gradual compression phase that selects most relevant spectro-temporal patterns using wide-pass 2 d filtering. Our results show that the approach significantly outperforms alternative waveform-based models on both noisy and spontaneous conversational speech (24% and 11% relative error reduction, respectively). Moreover, this study provides empirical evidence that learning directly from the waveform domain could be more effective than learning using hand-crafted features.

Raw waveform acoustic modelling has recently gained interest due to neural networks' ability to learn feature extraction, and the potential for finding better representations for a given scenario than hand-crafted features. SincNet has been proposed to reduce the number of parameters required in raw-waveform modelling, by restricting the filter functions, rather than having to learn every tap of each filter. We study the adaptation of the SincNet filter parameters from adults' to children's speech, and show that the parameterisation of the SincNet layer is well suited for adaptation in practice: we can efficiently adapt with a very small number of parameters, producing error rates comparable to techniques using orders of magnitude more parameters.

Speaker adaptive training (SAT) of neural network acoustic models learns models in a way that makes them more suitable for adaptation to test conditions. Conventionally, model-based speaker adaptive training is performed by having a set of speaker dependent parameters that are jointly optimised with speaker independent parameters in order to remove speaker variation. However, this does not scale well if all neural network weights are to be adapted to the speaker. In this paper we formulate speaker adaptive training as a meta-learning task, in which an adaptation process using gradient descent is encoded directly into the training of the model. We compare our approach with test-only adaptation of a standard baseline model and a SAT-LHUC model with a learned speaker adaptation schedule and demonstrate that the meta-learning approach achieves comparable results.

In this work, we investigate the use of embeddings for speaker-adaptive training of DNNs (DNN-SAT) focusing on a small amount of adaptation data per speaker. DNN-SAT can be viewed as learning a mapping from each embedding to transformation parameters that are applied to the shared parameters of the DNN. We investigate different approaches to applying these transformations, and find that with a good training strategy, a multi-layer adaptation network applied to all hidden layers is no more effective than a single linear layer acting on the embeddings to transform the input features. In the second part of our work, we evaluate different embed-dings (i-vectors, x-vectors and deep CNN embeddings) in an additional speaker recognition task in order to gain insight into what should characterize an embedding for DNN-SAT. We find the performance for speaker recognition of a given representation is not correlated with its ASR performance; in fact, ability to capture more speech attributes than just speaker identity was the most important characteristic of the embed-dings for efficient DNN-SAT ASR. Our best models achieved relative WER gains of 4% and 9% over DNN baselines using speaker-level cepstral mean normalisation (CMN), and a fully speaker-independent model, respectively.

DNNs play a major role in the state-of-the-art ASR systems. They can be used for extracting features and building probabilistic models for acoustic and language modelling. Despite their huge practical success, the level of theoretical understanding has remained shallow. This paper investigates DNNs from a statistical standpoint. In particular, the effect of activation functions on the distribution of the pre-activations and activations is investigated and discussed from both analytic and empirical viewpoints. This study, among others, shows that the pre-activation density in the bottleneck layer can be well fitted with a diagonal GMM with a few Gaussians and how and why the ReLU activation function promotes sparsity. Motivated by the statistical properties of the pre-activations, the usefulness of statistical normalisation of bottleneck features was also investigated. To this end, methods such as mean(-variance) normalisation, Gaussianisation, and histogram equalisation (HEQ) were employed and up to 2% (absolute) WER reduction achieved in the Aurora-4 task.

The usual attention mechanisms used for encoder-decoder models do not constrain the relationship between input and output sequences to be monotonic. To address this we explore windowed attention mechanisms which restrict attention to a block of source hidden states. Rule-based windowing restricts attention to a (typically large) fixed-length window. The performance of such methods is poor if the window size is small. In this paper, we propose a fully-trainable windowed attention and provide a detailed analysis on the factors which affect the performance of such an attention mechanism. Compared to the rule-based window methods, the learned window size is significantly smaller yet the model's performance is competitive. On the TIMIT corpus this approach has resulted in a 17% (relative) performance improvement over the traditional attention model. Our model also yields comparable accuracies to the joint CTC-attention model on the Wall Street Journal corpus.

Speech recognition models are highly susceptible to mismatch in the acoustic and language domains between the training and the evaluation data. For low resource languages, it is difficult to obtain transcribed speech for target domains, while untranscribed data can be collected with minimal effort. Recently, a method applying lattice-free maximum mutual information (LF-MMI) to untranscribed data has been found to be effective for semi-supervised training. However, weaker initial models and domain mismatch can result in high deletion rates for the semi-supervised model. Therefore, we propose a method to force the base model to overgenerate possible transcriptions, relying on the ability of LF-MMI to deal with uncertainty. On data from the IARPA MATERIAL programme, our new semi-supervised method outperforms the standard semisupervised method, yielding significant gains when adapting for mismatched bandwidth and domain.

In the broadcast domain there is an abundance of related text data and partial transcriptions, such as closed captions and subtitles. This text data can be used for lightly supervised training, in which text matching the audio is selected using an existing speech recognition model. Current approaches to light supervision typically filter the data based on matching error rates between the transcriptions and biased decoding hypotheses. In contrast, semi-supervised training does not require matching text data, instead generating a hypothesis using a background language model. State-of-the-art semi-supervised training uses lattice-based supervision with the lattice-free MMI (LF-MMI) objective function. We propose a technique to combine inaccurate transcriptions with the lattices generated for semi-supervised training, thus preserving uncertainty in the lattice where appropriate. We demonstrate that this combined approach reduces the expected error rates over the lattices, and reduces the word error rate (WER) on a broadcast task.

We investigate the problem of direct waveform modelling using parametric kernel-based filters in a convolutional neural network (CNN) framework, building on SincNet, a CNN employing the cardinal sine (sinc) function to implement learnable bandpass filters. To this end, the general problem of learning a filterbank consisting of modulated kernel-based baseband filters is studied. Compared to standard CNNs, such models have fewer parameters, learn faster, and require less training data. They are also more amenable to human interpretation, paving the way to embedding some perceptual prior knowledge in the architecture. We have investigated the replacement of the rectangular filters of SincNet with triangular, gammatone and Gaussian filters, resulting in higher model flexibility and a reduction to the phone error rate. We also explore the properties of the learned filters learned for TIMIT phone recognition from both perceptual and statistical standpoints. We find that the filters in the first layer, which directly operate on the waveform, are in accord with the prior knowledge utilised in designing and engineering standard filters such as mel-scale triangular filters. That is, the networks learn to pay more attention to perceptually significant spectral neighbourhoods where the data centroid is located, and the variance and Shannon entropy are highest.

Jointly optimised attention-based encoder-decoder models have yielded impressive speech recognition results. The recurrent neural network (RNN) encoder is a key component in such models – it learns the hidden representations of the inputs.However, it is difficult for RNNs to model the long sequences characteristic of speech recognition. To address this, subsampling between stacked recurrent layers of the encoder is commonly employed. This method reduces the length of the input sequence and leads to gains in accuracy. However, static subsampling may both include redundant information and miss relevant information. We propose using a dynamic subsampling RNN (dsRNN) encoder. Unlike a statically subsampled RNN encoder, the dsRNN encoder can learn to skip redundant frames. Furthermore, the skip ratio may vary at different stages of training, thus allowing the encoder to learn the most relevant information for each epoch. Although the dsRNN is unidirectional, it yields lower phone error rates (PERs) than a bidirectional RNN on TIMIT. The dsRNN encoder has a 16.8% PER on the TIMIT test set, a considerable improvement over static subsampling methods used with unidirectional and bidirectional RNN encoders (23.5% and 20.4% PER respectively).

The performance of automatic speech recognition systems can be improved by adapting an acoustic model to compensate for the mismatch between training and testing conditions, for example by adapting to unseen speakers. The success of speaker adaptation methods relies on selecting weights that are suitable for adaptation and using good adaptation schedules to update these weights in order not to overfit to the adaptation data. In this paper we investigate a principled way of adapting all the weights of the acoustic model using a meta-learning. We show that the meta-learner can learn to perform supervised and unsupervised speaker adaptation and that it outperforms a strong baseline adapting LHUC parameters when adapting a DNN AM with 1.5M parameters. We also report initial experiments on adapting TDNN AMs, where the meta-learner achieves comparable performance with LHUC.

We explore why deep convolutional neural networks (CNNs) with small two-dimensional kernels, primarily used for modeling spatial relations in images, are also effective in speech recognition. We analyze the representations learned by deep CNNs and compare them with deep neural network (DNN) representations and i-vectors, in the context of acoustic model adaptation. To explore whether interpretable information can be decoded from the learned representations we evaluate their ability to discriminate between speakers, acoustic conditions, noise type, and gender using the Aurora-4 dataset. We extract both whole model embeddings (to capture the information learned across the whole network) and layer-specific embeddings which enable understanding of the flow of information across the network. We also use learned representations as the additional input for a time-delay neural network (TDNN) for the Aurora-4 and MGB-3 English datasets. We find that deep CNN embeddings outperform DNN embeddings for acoustic model adaptation and auxiliary features based on deep CNN embeddings result in similar word error rates to i-vectors.

We study the problem of evaluating automatic speech recognition (ASR) systems that target dialectal speech input. A major challenge in this case is that the orthography of dialects is typically not standardized. From an ASR evaluation perspective, this means that there is no clear gold standard for the expected output, and several possible outputs could be considered correct according to different human annotators, which makes standard word error rate (WER) inadequate as an evaluation metric. Such a situation is typical for machine translation (MT), and thus we borrow ideas from an MT evaluation metric, namely TERp, an extension of translation error rate which is closely-related to WER. In particular, in the process of comparing a hypothesis to a reference, we make use of spelling variants for words and phrases, which we mine from Twitter in an unsupervised fashion. Our experiments with evaluating ASR output for Egyptian Arabic, and further manual analysis, show that the resulting WERd (i.e., WER for dialects) metric, a variant of TERp, is more adequate than WER for evaluating dialectal ASR.

Very deep convolutional neural networks (VDCNNs) have been successfully used in computer vision. More recently VDCNNs have been applied to speech recognition, using architectures adopted from computer vision. In this paper, we experimentally analyse the role of the components in VDCNN architectures for robust speech recognition. We have proposed a number of simplified VDCNN architectures, taking into account the use of fully-connected layers and down-sampling approaches. We have investigated three ways to down-sample feature maps: max-pooling, average-pooling, and convolution with increased stride. Our proposed model consisting solely of convolutional (conv) layers, and without any fully-connected layers, achieves a lower word error rate on Aurora 4 compared to other VDCNN architectures typically used in speech recognition. We have also extended our experiments to the MGB-3 task of multi-genre broadcast recognition using BBC TV recordings. The MGB-3 results indicate that the same architecture achieves the best result among our VDCNNs on this task as well.

A broadcast news stream consists of a number of stories and it is an important task to find the boundaries of stories automatically in news analysis. We capture the topic structure using a hierarchical model based on a Recurrent Neural Network (RNN) sentence modeling layer and a bidirectional Long Short-Term Memory (LSTM) topic modeling layer, with a fusion of acoustic and lexical features. Both features are accumulated with RNNs and trained jointly within the model to be fused at the sentence level. We conduct experiments on the topic detection and tracking (TDT4) task comparing combinations of two modalities trained with limited amount of parallel data. Further we utilize additional sufficient text data for training to polish our model. Experimental results indicate that the hierarchical RNN topic modeling takes advantage of the fusion scheme, especially with additional text training data, with a higher F1-measure compared to conventional state-of-the-art methods.

In this paper we present an extension of our previously described neural machine translation based system for punctuated transcription. This extension allows the system to map from per frame acoustic features to word level representations by replacing the traditional encoder in the encoder-decoder architecture with a hierarchical encoder. Furthermore, we show that a system combining lexical and acoustic features significantly outperforms systems using only a single source of features on all measured punctuation marks. The combination of lexical and acoustic features achieves a significant improvement in F-Measure of 1.5 absolute over the purely lexical neural machine translation based system.

Adapting acoustic models jointly to both speaker and environment has been shown to be effective. In many realistic scenarios, however, either the speaker or environment at test time might be unknown, or there may be insufficient data to learn a joint transform. Generating independent speaker and environment transforms improves the match of an acoustic model to unseen combinations. Using i-vectors, we demonstrate that it is possible to factorise speaker or environment information using multi-condition training with neural networks. Specifically, we extract bottleneck features from networks trained to classify either speakers or environments. We perform experiments on the Wall Street Journal corpus combined with environment noise from the Diverse Environments Multichannel Acoustic Noise Database. Using the factorised i-vectors we show improvements in word error rates on perturbed versions of the eval92 and dev93 test sets, both when one factor is missing and when the factors are seen but not in the desired combination.

A broadcast news stream consists of a number of stories and each story consists of several sentences. We capture this structure using a hierarchical model based on a word-level Recurrent Neural Network (RNN) sentence modeling layer and a sentence-level bidirectional Long Short-Term Memory (LSTM) topic modeling layer. First, the word-level RNN layer extracts a vector embedding the sentence information from the given transcribed lexical tokens of each sentence. These sentence embedding vectors are fed into a bidirectional LSTM that models the sentence and topic transitions. A topic posterior for each sentence is estimated discriminatively and a Hidden Markov model (HMM) follows to decode the story sequence and identify story boundaries. Experiments on the topic detection and tracking (TDT2) task indicate that the hierarchical RNN topic modeling achieves the best story segmentation performance with a higher F1-measure compared to conventional state-of-the-art methods. We also compare variations of our model to infer the optimal structure for the story segmentation task.

This paper investigates the use of multitask learning to improve context-dependent deep neural network (DNN) acoustic models. The use of hybrid DNN systems with clustered triphone targets is now standard in automatic speech recognition. However, we suggest that using a single set of DNN targets in this manner may not be the most effective choice, since the targets are the result of a somewhat arbitrary clustering process that may not be optimal for discrimination. We propose to remedy this problem through the addition of secondary tasks predicting alternative content-dependent or context-independent targets. We present a comprehensive set of experiments on a lecture recognition task showing that DNNs trained through multitask learning in this manner give consistently improved performance compared to standard hybrid DNNs. The technique is evaluated across a range of data and output sizes. Improvements are seen when training uses the cross entropy criterion and also when sequence training is applied.

We investigate different approaches for dialect identification in Arabic broadcast speech, using phonetic, lexical features obtained from a speech recognition system, and acoustic features using the i-vector framework. We studied both generative and discriminate classifiers, and we combined these features using a multi-class Support Vector Machine (SVM). We validated our results on an Arabic/English language identification task, with an accuracy of 100%. We used these features in a binary classifier to discriminate between Modern Standard Arabic (MSA) and Dialectal Arabic, with an accuracy of 100%. We further report results using the proposed method to discriminate between the five most widely used dialects of Arabic: namely Egyptian, Gulf, Levantine, North African, and MSA, with an accuracy of 52%. We discuss dialect identification errors in the context of dialect code-switching between Dialectal Arabic and MSA, and compare the error pattern between manually labeled data, and the output from our classifier. We also release the train and test data as standard corpus for dialect identification.

Children's speech poses challenges to speech recognition due to strong age-dependent anatomical variations and a lack of large, publicly-available corpora. In this paper we explore data augmentation for children's speech recognition using stochastic feature mapping (SFM) to transform out-of-domain adult data for both GMM-based and DNN-based acoustic models. We performed experiments on the English PF-STAR corpus, augmenting using WSJCAM0 and ABI. Our experimental results indicate that a DNN acoustic model for childrens speech can make use of adult data, and that out-of-domain SFM is more accurate than in-domain SFM.

Recurrent neural network language models (RNNLMs) have been shown to consistently improve Word Error Rates (WERs) of large vocabulary speech recognition systems employing n-gram LMs. In this paper we investigate supervised and unsupervised discriminative adaptation of RNNLMs in a broadcast transcription task to target domains defined by either genre or show. We have explored two approaches based on (1) scaling forward-propagated hidden activations (Learning Hidden Unit Contributions (LHUC) technique) and (2) direct fine-tuning of the parameters of the whole RNNLM. To investigate the effectiveness of the proposed methods we carry out experiments on multi-genre broadcast (MGB) data following the MGB-2015 challenge protocol. We observe small but significant improvements in WER compared to a strong unadapted RNNLM model.

This paper describes the Arabic Multi-Genre Broadcast (MGB-2) Challenge for SLT-2016. Unlike last year's English MGB Challenge, which focused on recognition of diverse TV genres, this year, the challenge has an emphasis on handling the diversity in dialect in Arabic speech. Audio data comes from 19 distinct programmes from the Aljazeera Arabic TV channel between March 2005 and December 2015. Programmes are split into three groups: conversations, interviews, and reports. A total of 1,200 hours have been released with lightly supervised transcriptions for the acoustic modelling. For language modelling, we made available over 110M words crawled from Aljazeera Arabic website Aljazeera.net for a 10 year duration 2000-2011. Two lexicons have been provided, one phoneme based and one grapheme based. Finally, two tasks were proposed for this year's challenge: standard speech transcription, and word alignment. This paper describes the task data and evaluation process used in the MGB challenge, and summarises the results obtained.

In this paper we investigate the punctuated transcription of multi-genre broadcast media. We examine four systems, three of which are based on lexical features, the fourth of which uses acoustic features by integrating punctuation into the speech recognition acoustic models. We also explore the combination of these component systems using voting and log-linear interpolation. We performed experiments on the English language MGB Challenge data, which comprises about 1,600h of BBC television recordings. Our results indicate that a lexical system, based on a neural machine translation approach is significantly better than other systems achieving an F-Measure of 62.6% on reference text, with a relative degradation of 19% on ASR output. Our analysis of the results in terms of specific punctuation indicated that using longer context improves the prediction of question marks and acoustic information improves prediction of exclamation marks. Finally, we show that even though the systems are complementary, their straightforward combination does not yield better F-measures than a single system using neural machine translation.

We describe our system for alignment of broadcast media captions in the 2015 MGB Challenge. A precise time alignment of previously-generated subtitles to media data is important in the process of caption generation by broadcasters. However, this task is challenging due to the highly diverse, often noisy content of the audio, and because the subtitles are frequently not a verbatim representation of the actual words spoken. Our system employs a two-pass approach with appropriately constrained weighted finite state transducers (WFSTs) to enable good alignment even when the audio quality would be challenging for conventional ASR. The system achieves an f-score of 0.8965 on the MGB Challenge development set.

Languages with no standard orthographic representation faces a challenge to evaluate the output from Automatic Speech Recognition (ASR). Since the reference transcription text can vary widely from one user to another. We propose an innovative approach for evaluating speech recognition using Multi-References. For each recognized speech segments, we ask five different users to transcribe the speech. We combine the alignment for the multiple references, and use the combined alignment to report a modified version of Word Error Rate (WER). This approach is in favor of accepting a recognized word if any of the references typed it in the same form. Results are reported using two Dialectal Arabic (DA) as a language with no standard orthographic; Egyptian, and North African speech. The average WER for the five references individually is 71.4%, and 80.1% respectively. When considering all references combined, the Multi-References MR-WER was found to be 39.7%, and 45.9% respectively.

This paper describes the Multi-Genre Broadcast (MGB) Challenge at ASRU 2015, an evaluation focused on speech recognition, speaker diarization, and "lightly supervised" alignment of BBC TV recordings. The challenge training data covered the whole range of seven weeks BBC TV output across four channels, resulting in about 1,600 hours of broadcast audio. In addition several hundred million words of BBC subtitle text was provided for language modelling. A novel aspect of the evaluation was the exploration of speech recognition and speaker diarization in a longitudinal setting — i.e. recognition of several episodes of the same show, and speaker diarization across these episodes, linking speakers. The longitudinal tasks also offered the opportunity for systems to make use of supplied metadata including show title, genre tag, and date/time of transmission. This paper describes the task data and evaluation process used in the MGB challenge, and summarises the results obtained.

The use of context-dependent targets has become standard in hybrid DNN systems for automatic speech recognition. However, we argue that despite the use of state-tying, optimising to context-dependent targets can lead to over-fitting, and that discriminating between arbitrary tied context-dependent targets may not be optimal. We propose a multitask learning method where the network jointly predicts context-dependent and monophone targets. We evaluate the method on a large-vocabulary lecture recognition task and show that it yields relative improvements of 3-10% over baseline systems.

We have previously found that context-dependent DNN models for automatic speech recognition can be improved with the use of monophone targets as a secondary task for the network. This paper asks whether the improvements derive from the regularising effect of having a much small number of monophone outputs – compared to the typical number of tied states – or from the use of targets that are not tied to an arbitrary stateclustering. We investigate the use of factorised targets for left and right context, and targets motivated by articulatory properties of the phonemes. We present results on a large-vocabulary lecture recognition task. Although the regularising effect of monophones seems to be important, all schemes give substantial improvements over the baseline single task system, even though the cardinality of the outputs is relatively high.

The phenomenon of reported speech -- whereby we quote the words, thoughts and opinions of others, or recount past dialogue -- is widespread in conversational speech. Detecting such quotations automatically has numerous applications: for example, in enhancing automatic transcription or spoken language understanding applications. However, the task is challenging, not least because lexical cues of quotations are frequently ambiguous or not present in spoken language. The aim of this paper is to identify potential prosodic cues of reported speech which could be used, along with the lexical ones, to automatically detect quotations and ascribe them to their rightful source, that is reconstructing their attribution relations. In order to do so we analyze SARC, a small corpus of telephone conversations that we have annotated with attribution relations. The results of the statistical analysis performed on the data show how variations in pitch, intensity, and timing features can be exploited as cues of quotations. Furthermore, we build a SVM classifier which integrates lexical and prosodic cues to automatically detect quotations in speech that performs significantly better than chance.

In previous work we have introduced a multi-task training tech-nique for neural network acoustic modelling, in which context-dependent and context-independent targets are jointly learned. In this paper, we extend the approach by structuring the out-put layer such that the context-dependent outputs are depen-dent on the context-independent outputs, thus using the context-independent predictions at run-time. We have also investigated the applicability of this idea to unsupervised speaker adapta-tion as an approach to overcome the data sparsity issues that comes to the fore when estimating systems with a large num-ber of context-dependent states, when data is limited. We have experimented with various amounts of training material (from 10 to 300 hours) and find the proposed techniques are particu-larly well suited to data-constrained conditions allowing to bet-ter utilise large context-dependent state-clustered trees. Exper-imental results are reported for large vocabulary speech recog-nition using the Switchboard and TED corpora. Index Terms: multitask learning, structured output layer, adap-tation, deep neural networks

Posterior-based or bottleneck features derived from neural net-works trained on out-of-domain data may be successfully ap-plied to improve speech recognition performance when data is scarce for the target domain or language. In this paper we com-bine this approach with the use of a hierarchical deep neural net-work (DNN) network structure – which we term a multi-level adaptive network (MLAN) – and the use of multitask learning. We have applied the technique to cross-lingual speech recog-nition experiments on recordings of TED talks and European Parliament sessions in English (source language) and German (target language). We demonstrate that the proposed method can lead to improvements over standard methods, even when the quantity of training data for the target language is relatively high. When the complete method is applied, we achieve rela-tiveWER reductions of around 13 % compared to a monolingual hybrid DNN baseline. Index Terms: deep neural network, multilevel adaptive net-

Speech segmentation is the problem of finding the end points of a speech utterance for passing to an automatic speech recogni-tion (ASR) system. The quality of this segmentation can have a large impact on the accuracy of the ASR system; in this pa-per we demonstrate that it can have an even larger impact on downstream natural language processing tasks – in this case, machine translation. We develop a novel semi-Markov model which allows the segmentation of audio streams into speech ut-terances which are optimised for the desired distribution of sen-tence lengths for the target domain. We compare this with exist-ing state-of-the-art methods and show that it is able to achieve not only improved ASR performance, but also to yield signifi-cant benefits to a speech translation task. Index Terms: speech activity detection, speech segmentation, machine translation, speech recognition

In this paper we investigate the use of Multi-level adaptive networks (MLAN) to incorporate out-of-domain data when training large vocabulary speech recognition systems. In a set of experiments on multi-genre broadcast data and on TED lecture recordings we present results using of out-of-domain features in a hybrid DNN system and explore tandem systems using a variety of input acoustic features. Our experiments indicate using the MLAN approach in both hybrid and tandem systems results in consistent reductions in word error rate of 5-10% relative.

Standard automatic speech recognition (ASR) systems use phonemes as subword units. Thus, one of the primary resource required to build a good ASR system is a well developed phoneme pronunciation lexicon. However, under-resourced languages typically lack such lexical resources. In this paper, we investigate recently proposed grapheme-based ASR in the framework of Kullback-Leibler divergence based hidden Markov model (KL-HMM) for underresourced languages, particularly Scottish Gaelic which has no lexical resources. More specifically, we study the use of grapheme and multilingual phoneme class conditional probabilities (posterior features) as feature observations in KL-HMM. ASR studies conducted show that the proposed approach yields better system compared to the conventional HMM/GMM approach using cepstral features. Furthermore, grapheme posterior features estimated using both auxiliary data and Gaelic data yield the best system.

This paper presents a new system for automatic transcription of lectures. The system combines a number of novel features, including deep neural network acoustic models using multi-level adaptive networks to incorporate out-of-domain information, and factored recurrent neural network language models. We demonstrate that the system achieves large improvements on the TED lecture transcription task from the 2012 IWSLT evaluation - our results are currently the best reported on this task, showing an relative WER reduction of more than 16% compared to the closest competing system from the evaluation.

Recently there has been increasing interest in ways of using out-of-domain (OOD) data to improve automatic speech recognition performance in domains where only limited data is available. This paper focuses on one such domain, namely that of disordered speech for which only very small databases exist, but where normal speech can be considered OOD. Standard approaches for handling small data domains use adaptation from OOD models into the target domain, but here we investigate an alternative approach with its focus on the feature extraction stage: OOD data is used to train feature-generating deep belief neural networks. Using AMI meeting and TED talk datasets, we investigate various tandem-based speaker independent systems as well as maximum a posteriori adapted speaker dependent systems. Results on the UAspeech isolated word task of disordered speech are very promising with our overall best system (using a combination of AMI and TED data) giving a correctness of 62.5 an increase of 15% on previously best published results based on conventional model adaptation. We show that the relative benefit of using OOD data varies considerably from speaker to speaker and is only loosely correlated with the severity of a speaker's impairments.

This paper introduces a method for lightly supervised discriminative training using MMI to improve the alignment of speech and text data for use in training HMM-based TTS systems for low-resource languages. In TTS applications, due to the use of long-span contexts, it is important to select training utterances which have wholly correct transcriptions. In a low-resource setting, when using poorly trained grapheme models, we show that the use of MMI discriminative training at the grapheme-level enables us to increase the amount of correctly aligned data by 40 while maintaining a 7% sentence error rate and 0.8% word error rate. We present the procedure for lightly supervised discriminative training with regard to the objective of minimising sentence error rate.

We describe our work in building a spoken language interface for a tutorial dialogue system. Our goal is to allow natural, unrestricted student interaction with the computer tutor, which has been shown to improve the student’s learning gain, but presents challenges for speech recognition and spoken language understanding. We discuss the choice of system components and present the results of development experiments in both acoustic and language modelling for speech recognition in this domain. Index Terms: spoken dialogue system, speech recognition, computer tutoring, adaptation

We describe our work on developing a speech recognition system for multi-genre media archives. The high diversity of the data makes this a challenging recognition task, which may benefit from systems trained on a combination of in-domain and out-of-domain data. Working with tandem HMMs, we present Multi-level Adaptive Networks (MLAN), a novel technique for incorporating information from out-of-domain posterior features using deep neural networks. We show that it provides a substantial reduction in WER over other systems, with relative WER reductions of 15% over a PLP baseline, 9% over in-domain tandem features and 8% over the best out-of-domain tandem features.

This paper introduces a method for automatic alignment of speech data with unsynchronised, imperfect transcripts, for a domain where no initial acoustic models are available. Using grapheme-based acoustic models, word skip networks and orthographic speech transcripts, we are able to harvest 55% of the speech with a 93% utterance-level accuracy and 99% word accuracy for the produced transcriptions. The work is based on the assumption that there is a high degree of correspondence between the speech and text, and that a full transcription of all of the speech is not required. The method is language independent and the only prior knowledge and resources required are the speech and text transcripts, and a few minor user interventions.

A major challenge faced by a spoken term detection (STD) system is the detection of out-of-vocabulary (OOV) terms. Although a subword-based STD system is able to detect OOV terms, performance reduction is always observed compared to in-vocabulary terms. One challenge that OOV terms bring to STD is the pronunciation uncertainty. A commonly used approach to address this problem is a soft matching procedure, and the other is the stochastic pronunciation modelling (SPM) proposed by the authors. In this paper we compare these two approaches, and combine them using a discriminative decision strategy. Experimental results demonstrated that SPM and soft match are highly complementary, and their combination gives significant performance improvement to OOV term detection.

We investigate the use of full covariance Gaussians for large-vocabulary speech recognition. The large number of parameters gives high modelling power, but when training data is limited, the standard sample covariance matrix is often poorly conditioned, and has high variance. We explain how these problems may be solved by the use of a diagonal covariance smoothing prior, and relate this to the shrinkage estimator, for which the optimal shrinkage parameter may itself be estimated from the training data. We also compare the use of generatively and discriminatively trained priors. Results are presented on a large vocabulary conversational telephone speech recognition task.

Within a spoken term detection (STD) system, the decision maker plays an important role in retrieving reliable detections. Most of the state-of-the-art STD systems make decisions based on a confidence measure that is term-independent, which poses a serious problem for out-of-vocabulary (OOV) term detection. In this paper, we study a term-dependent confidence measure based on confidence normalisation and discriminative modelling, particularly focusing on its remarkable effectiveness for detecting OOV terms. Experimental results indicate that the term-dependent confidence provides much more significant improvement for OOV terms than terms in-vocabulary. Index Terms: confidence estimation, spoken term detection, speech recognition

We consider the problem of parameter estimation in full-covariance Gaussian mixture systems for automatic speech recognition. Due to the high dimensionality of the acoustic feature vector, the standard sample covariance matrix has a high variance and is often poorly-conditioned when the amount of training data is limited. We explain how the use of a shrinkage estimator can solve these problems, and derive a formula for the optimal shrinkage intensity. We present results of experiments on a phone recognition task, showing that the estimator gives a performance improvement over a standard full-covariance system

We address the problem of learning the structure of Gaussian graphical models for use in automatic speech recognition, a means of controlling the form of the inverse covariance matrices of such systems. With particular focus on data sparsity issues, we implement a method for imposing graphical model structure on a Gaussian mixture system, using a convex optimisation technique to maximise a penalised likelihood expression. The results of initial experiments on a phone recognition task show a performance improvement over an equivalent full-covariance system. Index Terms: speech recognition, acoustic models, graphical models, precision matrix models