Peter Bell

Chao-Han Huck Yang, Taejin Park, Yuan Gong et al. · gatech

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.

Yuanchao Li, Yumnah Mohamied, Peter Bell et al.

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.

Ramon Sanabria, Nikolay Bogoychev, Nina Markl et al.

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 (https://groups.inf.ed.ac.uk/edacc/) under CC-BY-SA license.

Mathias Zinnen, Prathmesh Madhu, Ronak Kosti et al.

The Odeuropa Challenge on Olfactory Object Recognition aims to foster the development of object detection in the visual arts and to promote an olfactory perspective on digital heritage. Object detection in historical artworks is particularly challenging due to varying styles and artistic periods. Moreover, the task is complicated due to the particularity and historical variance of predefined target objects, which exhibit a large intra-class variance, and the long tail distribution of the dataset labels, with some objects having only very few training examples. These challenges should encourage participants to create innovative approaches using domain adaptation or few-shot learning. We provide a dataset of 2647 artworks annotated with 20 120 tightly fit bounding boxes that are split into a training and validation set (public). A test set containing 1140 artworks and 15 480 annotations is kept private for the challenge evaluation.

Christoph Minixhofer, Ondřej Klejch, Peter Bell

While modern Text-to-Speech (TTS) systems can produce natural-sounding speech, they remain unable to reproduce the full diversity found in natural speech data. We consider the distribution of all possible real speech samples that could be generated by these speakers alongside the distribution of all synthetic samples that could be generated for the same set of speakers, using a particular TTS system. We set out to quantify the distance between real and synthetic speech via a range of utterance-level statistics related to properties of the speaker, speech prosody and acoustic environment. Differences in the distribution of these statistics are evaluated using the Wasserstein distance. We reduce these distances by providing ground-truth values at generation time, and quantify the improvements to the overall distribution distance, approximated using an automatic speech recognition system. Our best system achieves a 10\% reduction in distribution distance.

Xiaoliang Wu, Peter Bell, Ajitha Rajan

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-Statistical Fault Localisation(SFL) and Causal. Additionally, we use an adapted version of Local Interpretable Model-Agnostic Explanations (LIME) 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, the baseline model of Sphinx, Deepspeech and 100 audio samples from the Commonvoice dataset.

Mathias Zinnen, Prathmesh Madhu, Peter Bell et al.

We investigate the effect of style and category similarity in multiple datasets used for object detection pretraining. We find that including an additional stage of object-detection pretraining can increase the detection performance considerably. While our experiments suggest that style similarities between pre-training and target datasets are less important than matching categories, further experiments are needed to verify this hypothesis.

Sarenne Wallbridge, Peter Bell, Catherine Lai

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.

Prathmesh Madhu, Tilman Marquart, Ronak Kosti et al.

Image compositions are helpful in the study of image structures and assist in discovering the semantics of the underlying scene portrayed across art forms and styles. With the digitization of artworks in recent years, thousands of images of a particular scene or narrative could potentially be linked together. However, manually linking this data with consistent objectiveness can be a highly challenging and time-consuming task. In this work, we present a novel approach called Image Composition Canvas (ICC++) to compare and retrieve images having similar compositional elements. ICC++ is an improvement over ICC specializing in generating low and high-level features (compositional elements) motivated by Max Imdahl's work. To this end, we present a rigorous quantitative and qualitative comparison of our approach with traditional and state-of-the-art (SOTA) methods showing that our proposed method outperforms all of them. In combination with deep features, our method outperforms the best deep learning-based method, opening the research direction for explainable machine learning for digital humanities. We will release the code and the data post-publication.

Yuanchao Li, Yuan Gong, Chao-Han Huck Yang et al.

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.

Yuanchao Li, Zixing Zhang, Jing Han et al.

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 Frechet 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.

Sung-Lin Yeh, Peter Bell, Hao Tang

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.

Christoph Minixhofer, Ondřej Klejch, Peter Bell

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.

Shree Harsha Bokkahalli Satish, Christoph Minixhofer, Maria Teleki et al.

Speech Large Language Models (SpeechLLMs) process spoken input directly, retaining cues such as accent and perceived gender that were previously removed in cascaded pipelines. This introduces speaker identity dependent variation in responses. We present a large-scale intersectional evaluation of accent and gender bias in three SpeechLLMs using 2,880 controlled interactions across six English accents and two gender presentations, keeping linguistic content constant through voice cloning. Using pointwise LLM-judge ratings, pairwise comparisons, and Best-Worst Scaling with human validation, we detect consistent disparities. Eastern European-accented speech receives lower helpfulness scores, particularly for female-presenting voices. The bias is implicit: responses remain polite but differ in helpfulness. While LLM judges capture the directional trend of these biases, human evaluators exhibit significantly higher sensitivity, uncovering sharper intersectional disparities.

Prathmesh Madhu, Tilman Marquart, Ronak Kosti et al.

Image compositions as a tool for analysis of artworks is of extreme significance for art historians. These compositions are useful in analyzing the interactions in an image to study artists and their artworks. Max Imdahl in his work called Ikonik, along with other prominent art historians of the 20th century, underlined the aesthetic and semantic importance of the structural composition of an image. Understanding underlying compositional structures within images is challenging and a time consuming task. Generating these structures automatically using computer vision techniques (1) can help art historians towards their sophisticated analysis by saving lot of time; providing an overview and access to huge image repositories and (2) also provide an important step towards an understanding of man made imagery by machines. In this work, we attempt to automate this process using the existing state of the art machine learning techniques, without involving any form of training. Our approach, inspired by Max Imdahl's pioneering work, focuses on two central themes of image composition: (a) detection of action regions and action lines of the artwork; and (b) pose-based segmentation of foreground and background. Currently, our approach works for artworks comprising of protagonists (persons) in an image. In order to validate our approach qualitatively and quantitatively, we conduct a user study involving experts and non-experts. The outcome of the study highly correlates with our approach and also demonstrates its domain-agnostic capability. We have open-sourced the code at https://github.com/image-compostion-canvas-group/image-compostion-canvas.

Prathmesh Madhu, Ronak Kosti, Lara Mührenberg et al.

In the field of Art History, images of artworks and their contexts are core to understanding the underlying semantic information. However, the highly complex and sophisticated representation of these artworks makes it difficult, even for the experts, to analyze the scene. From the computer vision perspective, the task of analyzing such artworks can be divided into sub-problems by taking a bottom-up approach. In this paper, we focus on the problem of recognizing the characters in Art History. From the iconography of $Annunciation$ $of$ $the$ $Lord$ (Figure 1), we consider the representation of the main protagonists, $Mary$ and $Gabriel$, across different artworks and styles. We investigate and present the findings of training a character classifier on features extracted from their face images. The limitations of this method, and the inherent ambiguity in the representation of $Gabriel$, motivated us to consider their bodies (a bigger context) to analyze in order to recognize the characters. Convolutional Neural Networks (CNN) trained on the bodies of $Mary$ and $Gabriel$ are able to learn person related features and ultimately improve the performance of character recognition. We introduce a new technique that generates more data with similar styles, effectively creating data in the similar domain. We present experiments and analysis on three different models and show that the model trained on domain related data gives the best performance for recognizing character. Additionally, we analyze the localized image regions for the network predictions. Code is open-sourced and available at https://github.com/prathmeshrmadhu/recognize_characters_art_history and the link to the published peer-reviewed article is https://dl.acm.org/citation.cfm?id=3357242.

Shree Harsha Bokkahalli Satish, Maria Teleki, Christoph Minixhofer et al.

SpeechLLMs process spoken language directly from audio, but accent and vocal identity cues can lead to biased behaviour. Current bias evaluations often miss how such bias manifests in end-to-end speech interactions and how users experience it. We distinguish quality-of-service disparities (e.g., off-topic or low-effort responses) from content-level bias in coherent outputs, and examine intersectional effects of accent and perceived gender. In this work, we explore a two-part evaluation approach: (1) a controlled test cohort spanning six accents and two gender presentations, analysed with judge-free prompt-response metrics, and (2) an interactive study design using voice conversion to let users experience identical content through different vocal identities. Across two studies (Interactive, N=24; Observational, N=19), we find that voice conversion increases trust and acceptability for benign responses and encourages perspective-taking, while automated analysis in search of quality-of-service disparities, reveals {accent x gender} disparities in alignment and verbosity across SpeechLLMs. These results highlight voice conversion for probing and experiencing intersectional voice bias while our evaluation suite provides richer bias evaluations for spoken conversational AI.

Jinzuomu Zhong, Yi Wang, Korin Richmond et al.

Discrete Speech Representation Tokens (DSRTs) have become a foundational component in speech generation. While prior work has extensively studied phonetic and speaker information in DSRTs, how accent information is encoded in DSRTs remains largely unexplored. In this paper, we present the first systematic investigation of accent information in DSRTs. We propose a unified evaluation framework that measures both accessibility of accent information via a novel Accent ABX task and recoverability via cross-accent Voice Conversion (VC) resynthesis. Using this framework, we analyse DSRTs derived from a variety of speech encoders. Our results reveal that accent information is substantially reduced when ASR supervision is used to fine-tune the encoder, but cannot be effectively disentangled from phonetic and speaker information through naive codebook size reduction. Based on these findings, we propose new content-only and content-accent DSRTs that significantly outperform existing designs in controllable accent generation. Our work highlights the importance of accent-aware evaluation and provides practical guidance for designing DSRTs for accent-controlled speech generation.

Dongge Han, Trevor McInroe, Adam Jelley et al.

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 individual user preferences. We introduce LLM-Personalize, a novel framework with an optimization pipeline designed to personalize LLM planners for household robotics. Our LLM-Personalize framework features an LLM planner that performs iterative planning in multi-room, partially-observable household scenarios, making use of a scene graph constructed with local observations. The generated plan consists of a sequence of high-level actions which are subsequently executed by a controller. Central to our approach is the optimization pipeline, which combines imitation learning and iterative self-training to personalize the LLM planner. In particular, the imitation learning phase performs initial LLM alignment from demonstrations, and bootstraps the model to facilitate effective iterative self-training, which further explores and aligns the model to user preferences. We evaluate LLM-Personalize on Housekeep, a challenging simulated real-world 3D benchmark for household rearrangements, and show that LLM-Personalize achieves more than a 30 percent increase in success rate over existing LLM planners, showcasing significantly improved alignment with human preferences. Project page: https://gdg94.github.io/projectllmpersonalize/.

Mathias Zinnen, Prathmesh Madhu, Inger Leemans et al.

Real-world applications of computer vision in the humanities require algorithms to be robust against artistic abstraction, peripheral objects, and subtle differences between fine-grained target classes. Existing datasets provide instance-level annotations on artworks but are generally biased towards the image centre and limited with regard to detailed object classes. The proposed ODOR dataset fills this gap, offering 38,116 object-level annotations across 4712 images, spanning an extensive set of 139 fine-grained categories. Conducting a statistical analysis, we showcase challenging dataset properties, such as a detailed set of categories, dense and overlapping objects, and spatial distribution over the whole image canvas. Furthermore, we provide an extensive baseline analysis for object detection models and highlight the challenging properties of the dataset through a set of secondary studies. Inspiring further research on artwork object detection and broader visual cultural heritage studies, the dataset challenges researchers to explore the intersection of object recognition and smell perception.

Christoph Minixhofer, Ondrej Klejch, Peter Bell

Evaluation of Text to Speech (TTS) systems is challenging and resource-intensive. Subjective metrics such as Mean Opinion Score (MOS) are not easily comparable between works. Objective metrics are frequently used, but rarely validated against subjective ones. Both kinds of metrics are challenged by recent TTS systems capable of producing synthetic speech indistinguishable from real speech. In this work, we introduce Text to Speech Distribution Score 2 (TTSDS2), a more robust and improved version of TTSDS. Across a range of domains and languages, it is the only one out of 16 compared metrics to correlate with a Spearman correlation above 0.50 for every domain and subjective score evaluated. We also release a range of resources for evaluating synthetic speech close to real speech: A dataset with over 11,000 subjective opinion score ratings; a pipeline for continually recreating a multilingual test dataset to avoid data leakage; and a continually updated benchmark for TTS in 14 languages.

Ondřej Klejch, William Lamb, Peter Bell

An effective approach to the development of ASR systems for low-resource languages is to fine-tune an existing multilingual end-to-end model. When the original model has been trained on large quantities of data from many languages, fine-tuning can be effective with limited training data, even when the language in question was not present in the original training data. The fine-tuning approach has been encouraged by the availability of public-domain E2E models and is widely believed to lead to state-of-the-art results. This paper, however, challenges that belief. We show that an approach combining hybrid HMMs with self-supervised models can yield substantially better performance with limited training data. This combination allows better utilisation of all available speech and text data through continued self-supervised pre-training and semi-supervised training. We benchmark our approach on Scottish Gaelic, achieving WER reductions of 32% relative over our best fine-tuned Whisper model.

Sarenne Wallbridge, Christoph Minixhofer, Catherine Lai et al.

People exploit the predictability of lexical structures during text comprehension. Though predictable structure is also present in speech, the degree to which prosody, e.g. intonation, tempo, and loudness, contributes to such structure independently of the lexical content is unclear. This study leverages self-supervised learning (SSL) to examine the temporal granularity of structures in the acoustic correlates of prosody. Representations from our proposed Masked Prosody Model can predict perceptual labels dependent on local information, such as word boundaries, but provide the most value for labels involving longer-term structures, like emotion recognition. Probing experiments across various perceptual labels show strong relative gains over untransformed pitch, energy, and voice activity features. Our results reveal the importance of SSL training objective timescale and highlight the value of complex SSL-encoded structures compared to more constrained classical structures.

Edward Storey, Naomi Harte, Peter Bell

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.

Christoph Minixhofer, Ondrej Klejch, Peter Bell

Synthetically generated speech has rapidly approached human levels of naturalness. However, the paradox remains that ASR systems, when trained on TTS output that is judged as natural by humans, continue to perform badly on real speech. In this work, we explore whether this phenomenon is due to the oversmoothing behaviour of models commonly used in TTS, with a particular focus on the behaviour of TTS-for-ASR as the amount of TTS training data is scaled up. We systematically compare Denoising Diffusion Probabilistic Models (DDPM) to Mean Squared Error (MSE) based models for TTS, when used for ASR model training. We test the scalability of the two approaches, varying both the number hours, and the number of different speakers. We find that for a given model size, DDPM can make better use of more data, and a more diverse set of speakers, than MSE models. We achieve the best reported ratio between real and synthetic speech WER to date (1.46), but also find that a large gap remains.

Yuanchao Li, Peter Bell, Catherine Lai

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.

Erfan Loweimi, Andrea Carmantini, Peter Bell et al.

In this paper, we analyse the error patterns of the raw waveform acoustic models in TIMIT's phone recognition task. Our analysis goes beyond the conventional phone error rate (PER) metric. We categorise the phones into three groups: {affricate, diphthong, fricative, nasal, plosive, semi-vowel, vowel, silence}, {consonant, vowel+, silence}, and {voiced, unvoiced, silence} and, compute the PER for each broad phonetic class in each category. We also construct a confusion matrix for each category using the substitution errors and compare the confusion patterns with those of the Filterbank and Wav2vec 2.0 systems. Our raw waveform acoustic models consists of parametric (Sinc2Net) or non-parametric CNNs and Bidirectional LSTMs, achieving down to 13.7%/15.2% PERs on TIMIT Dev/Test sets, outperforming reported PERs for raw waveform models in the literature. We also investigate the impact of transfer learning from WSJ on the phonetic error patterns and confusion matrices. It reduces the PER to 11.8%/13.7% on the Dev/Test sets.

Xiaoliang Wu, Peter Bell, Ajitha Rajan

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.

Yuanchao Li, Zeyu Zhao, Ondrej Klejch et al.

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, wav2vec2, 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.

Yaoting Wang, Yuanchao Li, Paul Pu Liang et al.

Fusing multiple modalities has proven effective for multimodal information processing. However, the incongruity between modalities poses a challenge for multimodal fusion, especially in affect recognition. In this study, we first analyze how the salient affective information in one modality can be affected by the other, and demonstrate that inter-modal incongruity exists latently in crossmodal attention. Based on this finding, we propose the Hierarchical Crossmodal Transformer with Dynamic Modality Gating (HCT-DMG), a lightweight incongruity-aware model, which dynamically chooses the primary modality in each training batch and reduces fusion times by leveraging the learned hierarchy in the latent space to alleviate incongruity. The experimental evaluation on five benchmark datasets: CMU-MOSI, CMU-MOSEI, and IEMOCAP (sentiment and emotion), where incongruity implicitly lies in hard samples, as well as UR-FUNNY (humour) and MUStaRD (sarcasm), where incongruity is common, verifies the efficacy of our approach, showing that HCT-DMG: 1) outperforms previous multimodal models with a reduced size of approximately 0.8M parameters; 2) recognizes hard samples where incongruity makes affect recognition difficult; 3) mitigates the incongruity at the latent level in crossmodal attention.

Christoph Minixhofer, Ondřej Klejch, Peter Bell

In this work, we unify several existing decoding strategies for punctuation prediction in one framework and introduce a novel strategy which utilises multiple predictions at each word across different windows. We show that significant improvements can be achieved by optimising these strategies after training a model, only leading to a potential increase in inference time, with no requirement for retraining. We further use our decoding strategy framework for the first comparison of tagging and classification approaches for punctuation prediction in a real-time setting. Our results show that a classification approach for punctuation prediction can be beneficial when little or no right-side context is available.

Ondrej Klejch, Electra Wallington, Peter Bell

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.

Yuanchao Li, Peter Bell, Catherine Lai

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.

Sarenne Wallbridge, Peter Bell, Catherine Lai

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.

David Wan, Chris Kedzie, Faisal Ladhak et al.

Typical ASR systems segment the input audio into utterances using purely acoustic information, which may not resemble the sentence-like units that are expected by conventional machine translation (MT) systems for Spoken Language Translation. In this work, we propose a model for correcting the acoustic segmentation of ASR models for low-resource languages to improve performance on downstream tasks. We propose the use of subtitles as a proxy dataset for correcting ASR acoustic segmentation, creating synthetic acoustic utterances by modeling common error modes. We train a neural tagging model for correcting ASR acoustic segmentation and show that it improves downstream performance on MT and audio-document cross-language information retrieval (CLIR).

Shucong Zhang, Cong-Thanh Do, Rama Doddipatla et al.

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.

Prathmesh Madhu, Angel Villar-Corrales, Ronak Kosti et al.

Human pose estimation (HPE) is a central part of understanding the visual narration and body movements of characters depicted in artwork collections, such as Greek vase paintings. Unfortunately, existing HPE methods do not generalise well across domains resulting in poorly recognized poses. Therefore, we propose a two step approach: (1) adapting a dataset of natural images of known person and pose annotations to the style of Greek vase paintings by means of image style-transfer. We introduce a perceptually-grounded style transfer training to enforce perceptual consistency. Then, we fine-tune the base model with this newly created dataset. We show that using style-transfer learning significantly improves the SOTA performance on unlabelled data by more than 6% mean average precision (mAP) as well as mean average recall (mAR). (2) To improve the already strong results further, we created a small dataset (ClassArch) consisting of ancient Greek vase paintings from the 6-5th century BCE with person and pose annotations. We show that fine-tuning on this data with a style-transferred model improves the performance further. In a thorough ablation study, we give a targeted analysis of the influence of style intensities, revealing that the model learns generic domain styles. Additionally, we provide a pose-based image retrieval to demonstrate the effectiveness of our method.

Shucong Zhang, Erfan Loweimi, Peter Bell et al.

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.

Shucong Zhang, Erfan Loweimi, Peter Bell et al.

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.

Chau Luu, Peter Bell, Steve Renals

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.

David Wan, Zhengping Jiang, Chris Kedzie et al.

In this work, we focus on improving ASR output segmentation in the context of low-resource language speech-to-text translation. ASR output segmentation is crucial, as ASR systems segment the input audio using purely acoustic information and are not guaranteed to output sentence-like segments. Since most MT systems expect sentences as input, feeding in longer unsegmented passages can lead to sub-optimal performance. We explore the feasibility of using datasets of subtitles from TV shows and movies to train better ASR segmentation models. We further incorporate part-of-speech (POS) tag and dependency label information (derived from the unsegmented ASR outputs) into our segmentation model. We show that this noisy syntactic information can improve model accuracy. We evaluate our models intrinsically on segmentation quality and extrinsically on downstream MT performance, as well as downstream tasks including cross-lingual information retrieval (CLIR) tasks and human relevance assessments. Our model shows improved performance on downstream tasks for Lithuanian and Bulgarian.

Peter Bell, Joachim Fainberg, Ondrej Klejch et al.

We present a structured overview of adaptation algorithms for neural network-based speech recognition, considering both hybrid hidden Markov model / neural network systems and end-to-end neural network systems, with a focus on speaker adaptation, domain adaptation, and accent adaptation. The overview characterizes adaptation algorithms as based on embeddings, model parameter adaptation, or data augmentation. We present a meta-analysis of the performance of speech recognition adaptation algorithms, based on relative error rate reductions as reported in the literature.

Shucong Zhang, Erfan Loweimi, Peter Bell et al.

Recently, self-attention models such as Transformers have given competitive results compared to recurrent neural network systems in speech recognition. The key factor for the outstanding performance of self-attention models is their ability to capture temporal relationships without being limited by the distance between two related events. However, we note that the range of the learned context progressively 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 still important for the upper self-attention layers in the encoder of Transformers? To investigate this, we replace these self-attention layers with feed forward layers. In our speech recognition experiments (Wall Street Journal and Switchboard), we indeed observe an interesting result: replacing the upper self-attention layers in the encoder with feed forward layers leads to no performance drop, and even minor gains. Our experiments offer insights to how self-attention layers process the speech signal, leading to the conclusion that the lower self-attention layers of the encoder encode a sufficiently wide range of inputs, hence learning further contextual information in the upper layers is unnecessary.

Chau Luu, Peter Bell, Steve Renals

Many recent works on deep speaker embeddings train their feature extraction networks on large classification tasks, distinguishing between all speakers in a training set. Empirically, this has been shown to produce speaker-discriminative embeddings, even for unseen speakers. However, it is not clear that this is the optimal means of training embeddings that generalize well. This work proposes two approaches to learning embeddings, based on the notion of dropping classes during training. We demonstrate that both approaches can yield performance gains in speaker verification tasks. The first proposed method, DropClass, works via periodically dropping a random subset of classes from the training data and the output layer throughout training, resulting in a feature extractor trained on many different classification tasks. Combined with an additive angular margin loss, this method can yield a 7.9% relative improvement in equal error rate (EER) over a strong baseline on VoxCeleb. The second proposed method, DropAdapt, is a means of adapting a trained model to a set of enrolment speakers in an unsupervised manner. This is performed by fine-tuning a model on only those classes which produce high probability predictions when the enrolment speakers are used as input, again also dropping the relevant rows from the output layer. This method yields a large 13.2% relative improvement in EER on VoxCeleb. The code for this paper has been made publicly available.

Joanna Rownicka, Peter Bell, Steve Renals

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.

Chau Luu, Peter Bell, Steve Renals

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.

Ondřej Klejch, Joachim Fainberg, Peter Bell et al.

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.

Joachim Fainberg, Ondřej Klejch, Erfan Loweimi et al.

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.

Joanna Rownicka, Peter Bell, Steve Renals

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 embeddings (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 embeddings 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.

Ondrej Klejch, Joachim Fainberg, Peter Bell et al.

Acoustic model adaptation to unseen test recordings aims to reduce the mismatch between training and testing conditions. Most adaptation schemes for neural network models require the use of an initial one-best transcription for the test data, generated by an unadapted model, in order to estimate the adaptation transform. It has been found that adaptation methods using discriminative objective functions - such as cross-entropy loss - often require careful regularisation to avoid over-fitting to errors in the one-best transcriptions. In this paper we solve this problem by performing discriminative adaptation using lattices obtained from a first pass decoding, an approach that can be readily integrated into the lattice-free maximum mutual information (LF-MMI) framework. We investigate this approach on three transcription tasks of varying difficulty: TED talks, multi-genre broadcast (MGB) and a low-resource language (Somali). We find that our proposed approach enables many more parameters to be adapted without over-fitting being observed, and is successful even when the initial transcription has a WER in excess of 50%.