Ewan Dunbar

Juliette Millet, Charlotte Caucheteux, Pierre Orhan et al.

Several deep neural networks have recently been shown to generate activations similar to those of the brain in response to the same input. These algorithms, however, remain largely implausible: they require (1) extraordinarily large amounts of data, (2) unobtainable supervised labels, (3) textual rather than raw sensory input, and / or (4) implausibly large memory (e.g. thousands of contextual words). These elements highlight the need to identify algorithms that, under these limitations, would suffice to account for both behavioral and brain responses. Focusing on the issue of speech processing, we here hypothesize that self-supervised algorithms trained on the raw waveform constitute a promising candidate. Specifically, we compare a recent self-supervised architecture, Wav2Vec 2.0, to the brain activity of 412 English, French, and Mandarin individuals recorded with functional Magnetic Resonance Imaging (fMRI), while they listened to ~1h of audio books. Our results are four-fold. First, we show that this algorithm learns brain-like representations with as little as 600 hours of unlabelled speech -- a quantity comparable to what infants can be exposed to during language acquisition. Second, its functional hierarchy aligns with the cortical hierarchy of speech processing. Third, different training regimes reveal a functional specialization akin to the cortex: Wav2Vec 2.0 learns sound-generic, speech-specific and language-specific representations similar to those of the prefrontal and temporal cortices. Fourth, we confirm the similarity of this specialization with the behavior of 386 additional participants. These elements, resulting from the largest neuroimaging benchmark to date, show how self-supervised learning can account for a rich organization of speech processing in the brain, and thus delineate a path to identify the laws of language acquisition which shape the human brain.

Tu Anh Nguyen, Maureen de Seyssel, Robin Algayres et al. · apple-ml

Word or word-fragment based Language Models (LM) are typically preferred over character-based ones in many downstream applications. This may not be surprising as words seem more linguistically relevant units than characters. Words provide at least two kinds of relevant information: boundary information and meaningful units. However, word boundary information may be absent or unreliable in the case of speech input (word boundaries are not marked explicitly in the speech stream). Here, we systematically compare LSTMs as a function of the input unit (character, phoneme, word, word part), with or without gold boundary information. We probe linguistic knowledge in the networks at the lexical, syntactic and semantic levels using three speech-adapted black box NLP psycholinguistically-inpired benchmarks (pWUGGY, pBLIMP, pSIMI). We find that the absence of boundaries costs between 2\% and 28\% in relative performance depending on the task. We show that gold boundaries can be replaced by automatically found ones obtained with an unsupervised segmentation algorithm, and that even modest segmentation performance gives a gain in performance on two of the three tasks compared to basic character/phone based models without boundary information.

Ewan Dunbar, Nicolas Hamilakis, Emmanuel Dupoux

Recent progress in self-supervised or unsupervised machine learning has opened the possibility of building a full speech processing system from raw audio without using any textual representations or expert labels such as phonemes, dictionaries or parse trees. The contribution of the Zero Resource Speech Challenge series since 2015 has been to break down this long-term objective into four well-defined tasks -- Acoustic Unit Discovery, Spoken Term Discovery, Discrete Resynthesis, and Spoken Language Modeling -- and introduce associated metrics and benchmarks enabling model comparison and cumulative progress. We present an overview of the six editions of this challenge series since 2015, discuss the lessons learned, and outline the areas which need more work or give puzzling results.

Juliette Millet, Ewan Dunbar

Self-supervised models for speech processing form representational spaces without using any external labels. Increasingly, they appear to be a feasible way of at least partially eliminating costly manual annotations, a problem of particular concern for low-resource languages. But what kind of representational spaces do these models construct? Human perception specializes to the sounds of listeners' native languages. Does the same thing happen in self-supervised models? We examine the representational spaces of three kinds of state-of-the-art self-supervised models: wav2vec 2.0, HuBERT and contrastive predictive coding (CPC), and compare them with the perceptual spaces of French-speaking and English-speaking human listeners, both globally and taking account of the behavioural differences between the two language groups. We show that the CPC model shows a small native language effect, but that wav2vec 2.0 and HuBERT seem to develop a universal speech perception space which is not language specific. A comparison against the predictions of supervised phone recognisers suggests that all three self-supervised models capture relatively fine-grained perceptual phenomena, while supervised models are better at capturing coarser, phone-level, effects of listeners' native language, on perception.

Mark Hallap, Emmanuel Dupoux, Ewan Dunbar

Unsupervised speech representations have taken off, with benchmarks (SUPERB, ZeroSpeech) demonstrating major progress on semi-supervised speech recognition, speech synthesis, and speech-only language modelling. Inspiration comes from the promise of ``discovering the phonemes'' of a language or a similar low-bitrate encoding. However, one of the critical properties of phoneme transcriptions is context-invariance: the phonetic context of a speech sound can have massive influence on the way it is pronounced, while the text remains stable. This is what allows tokens of the same word to have the same transcriptions -- key to language understanding. Current benchmarks do not measure context-invariance. We develop a new version of the ZeroSpeech ABX benchmark that measures context-invariance, and apply it to recent self-supervised representations. We demonstrate that the context-independence of representations is predictive of the stability of word-level representations. We suggest research concentrate on improving context-independence of self-supervised and unsupervised representations.

Juliette Millet, Ioana Chitoran, Ewan Dunbar

Our native language influences the way we perceive speech sounds, affecting our ability to discriminate non-native sounds. We compare two ideas about the influence of the native language on speech perception: the Perceptual Assimilation Model, which appeals to a mental classification of sounds into native phoneme categories, versus the idea that rich, fine-grained phonetic representations tuned to the statistics of the native language, are sufficient. We operationalize this idea using representations from two state-of-the-art speech models, a Dirichlet process Gaussian mixture model and the more recent wav2vec 2.0 model. We present a new, open dataset of French- and English-speaking participants' speech perception behaviour for 61 vowel sounds from six languages. We show that phoneme assimilation is a better predictor than fine-grained phonetic modelling, both for the discrimination behaviour as a whole, and for predicting differences in discriminability associated with differences in native language background. We also show that wav2vec 2.0, while not good at capturing the effects of native language on speech perception, is complementary to information about native phoneme assimilation, and provides a good model of low-level phonetic representations, supporting the idea that both categorical and fine-grained perception are used during speech perception.

Kuan-Po Huang, Chih-Kai Yang, Yu-Kuan Fu et al.

We introduce a new zero resource code-switched speech benchmark designed to directly assess the code-switching capabilities of self-supervised speech encoders. We showcase a baseline system of language modeling on discrete units to demonstrate how the code-switching abilities of speech encoders can be assessed in a zero-resource manner. Our experiments encompass a variety of well-known speech encoders, including Wav2vec 2.0, HuBERT, XLSR, etc. We examine the impact of pre-training languages and model size on benchmark performance. Notably, though our results demonstrate that speech encoders with multilingual pre-training, exemplified by XLSR, outperform monolingual variants (Wav2vec 2.0, HuBERT) in code-switching scenarios, there is still substantial room for improvement in their code-switching linguistic abilities.

Maxime Poli, Manel Khentout, Angelo Ortiz Tandazo et al.

We introduce DiscoPhon, a multilingual benchmark for evaluating unsupervised phoneme discovery from discrete speech units. DiscoPhon covers 6 dev and 6 test languages, chosen to span a wide range of phonemic contrasts. Given only 10 hours of speech in a previously unseen language, systems must produce discrete units that are mapped to a predefined phoneme inventory, through either a many-to-one or a one-to-one assignment. The resulting sequences are evaluated for unit quality, recognition and segmentation. We provide four pretrained multilingual HuBERT and SpidR baselines, and show that phonemic information is available enough in current models for derived units to correlate well with phonemes, though with variations across languages.

Sean Robertson, Gerald Penn, Ewan Dunbar

A long-standing question in automatic speech recognition research is how to attribute errors to the ability of a model to model the acoustics, versus its ability to leverage higher-order context (lexicon, morphology, syntax, semantics). We validate a novel approach which models error rates as a function of relative textual predictability, and yields a single number, $k$, which measures the effect of textual predictability on the recognizer. We use this method to demonstrate that a Wav2Vec 2.0-based model makes greater stronger use of textual context than a hybrid ASR model, in spite of not using an explicit language model, and also use it to shed light on recent results demonstrating poor performance of standard ASR systems on African-American English. We demonstrate that these mostly represent failures of acoustic--phonetic modelling. We show how this approach can be used straightforwardly in diagnosing and improving ASR.

Michael Ong, Sean Robertson, Leo Peckham et al.

We introduce the Faetar Automatic Speech Recognition Benchmark, a benchmark corpus designed to push the limits of current approaches to low-resource speech recognition. Faetar, a Franco-Provençal variety spoken primarily in Italy, has no standard orthography, has virtually no existing textual or speech resources other than what is included in the benchmark, and is quite different from other forms of Franco-Provençal. The corpus comes from field recordings, most of which are noisy, for which only 5 hrs have matching transcriptions, and for which forced alignment is of variable quality. The corpus contains an additional 20 hrs of unlabelled speech. We report baseline results from state-of-the-art multilingual speech foundation models with a best phone error rate of 30.4%, using a pipeline that continues pre-training on the foundation model using the unlabelled set.

Robin Huo, Ewan Dunbar

Self-supervised models for speech representation learning now see widespread use for their versatility and performance on downstream tasks, but the effect of model architecture on the linguistic information learned in their representations remains under-studied. This study investigates two such models, HuBERT and wav2vec 2.0, and minimally compares two of their architectural differences: training objective and iterative pseudo-label refinement through multiple training iterations. We find that differences in canonical correlation of hidden representations to word identity, phoneme identity, and speaker identity are explained by training iteration, not training objective. We suggest that future work investigate the reason for the effectiveness of iterative refinement in encoding linguistic information in self-supervised speech representations.

Leo Peckham, Michael Ong, Naomi Nagy et al.

We examine the role of transcription inconsistencies in the Faetar Automatic Speech Recognition benchmark, a challenging low-resource ASR benchmark. With the help of a small, hand-constructed lexicon, we conclude that find that, while inconsistencies do exist in the transcriptions, they are not the main challenge in the task. We also demonstrate that bigram word-based language modelling is of no added benefit, but that constraining decoding to a finite lexicon can be beneficial. The task remains extremely difficult.

Ewan Dunbar, Mathieu Bernard, Nicolas Hamilakis et al.

We present the Zero Resource Speech Challenge 2021, which asks participants to learn a language model directly from audio, without any text or labels. The challenge is based on the Libri-light dataset, which provides up to 60k hours of audio from English audio books without any associated text. We provide a pipeline baseline system consisting on an encoder based on contrastive predictive coding (CPC), a quantizer ($k$-means) and a standard language model (BERT or LSTM). The metrics evaluate the learned representations at the acoustic (ABX discrimination), lexical (spot-the-word), syntactic (acceptability judgment) and semantic levels (similarity judgment). We present an overview of the eight submitted systems from four groups and discuss the main results.

Louis Fournier, Ewan Dunbar

Many types of distributional word embeddings (weakly) encode linguistic regularities as directions (the difference between "jump" and "jumped" will be in a similar direction to that of "walk" and "walked," and so on). Several attempts have been made to explain this fact. We respond to Allen and Hospedales' recent (ICML, 2019) theoretical explanation, which claims that word2vec and GloVe will encode linguistic regularities whenever a specific relation of paraphrase holds between the four words involved in the regularity. We demonstrate that the explanation does not go through: the paraphrase relations needed under this explanation do not hold empirically.

Tu Anh Nguyen, Maureen de Seyssel, Patricia Rozé et al.

We introduce a new unsupervised task, spoken language modeling: the learning of linguistic representations from raw audio signals without any labels, along with the Zero Resource Speech Benchmark 2021: a suite of 4 black-box, zero-shot metrics probing for the quality of the learned models at 4 linguistic levels: phonetics, lexicon, syntax and semantics. We present the results and analyses of a composite baseline made of the concatenation of three unsupervised systems: self-supervised contrastive representation learning (CPC), clustering (k-means) and language modeling (LSTM or BERT). The language models learn on the basis of the pseudo-text derived from clustering the learned representations. This simple pipeline shows better than chance performance on all four metrics, demonstrating the feasibility of spoken language modeling from raw speech. It also yields worse performance compared to text-based 'topline' systems trained on the same data, delineating the space to be explored by more sophisticated end-to-end models.

Ewan Dunbar, Julien Karadayi, Mathieu Bernard et al.

We present the Zero Resource Speech Challenge 2020, which aims at learning speech representations from raw audio signals without any labels. It combines the data sets and metrics from two previous benchmarks (2017 and 2019) and features two tasks which tap into two levels of speech representation. The first task is to discover low bit-rate subword representations that optimize the quality of speech synthesis; the second one is to discover word-like units from unsegmented raw speech. We present the results of the twenty submitted models and discuss the implications of the main findings for unsupervised speech learning.

Juliette Millet, Ewan Dunbar

In this paper, we present a data set and methods to compare speech processing models and human behaviour on a phone discrimination task. We provide Perceptimatic, an open data set which consists of French and English speech stimuli, as well as the results of 91 English- and 93 French-speaking listeners. The stimuli test a wide range of French and English contrasts, and are extracted directly from corpora of natural running read speech, used for the 2017 Zero Resource Speech Challenge. We provide a method to compare humans' perceptual space with models' representational space, and we apply it to models previously submitted to the Challenge. We show that, unlike unsupervised models and supervised multilingual models, a standard supervised monolingual HMM-GMM phone recognition system, while good at discriminating phones, yields a representational space very different from that of human native listeners.

Louis Fournier, Emmanuel Dupoux, Ewan Dunbar

Vector space models of words have long been claimed to capture linguistic regularities as simple vector translations, but problems have been raised with this claim. We decompose and empirically analyze the classic arithmetic word analogy test, to motivate two new metrics that address the issues with the standard test, and which distinguish between class-wise offset concentration (similar directions between pairs of words drawn from different broad classes, such as France--London, China--Ottawa, ...) and pairing consistency (the existence of a regular transformation between correctly-matched pairs such as France:Paris::China:Beijing). We show that, while the standard analogy test is flawed, several popular word embeddings do nevertheless encode linguistic regularities.

Juliette Millet, Ewan Dunbar

We present the Perceptimatic English Benchmark, an open experimental benchmark for evaluating quantitative models of speech perception in English. The benchmark consists of ABX stimuli along with the responses of 91 American English-speaking listeners. The stimuli test discrimination of a large number of English and French phonemic contrasts. They are extracted directly from corpora of read speech, making them appropriate for evaluating statistical acoustic models (such as those used in automatic speech recognition) trained on typical speech data sets. We show that phone discrimination is correlated with several types of models, and give recommendations for researchers seeking easily calculated norms of acoustic distance on experimental stimuli. We show that DeepSpeech, a standard English speech recognizer, is more specialized on English phoneme discrimination than English listeners, and is poorly correlated with their behaviour, even though it yields a low error on the decision task given to humans.

Maud Parrot, Juliette Millet, Ewan Dunbar

Reconstruction of articulatory trajectories from the acoustic speech signal has been proposed for improving speech recognition and text-to-speech synthesis. However, to be useful in these settings, articulatory reconstruction must be speaker independent. Furthermore, as most research focuses on single, small datasets with few speakers, robust articulatory reconstrucion could profit from combining datasets. Standard evaluation measures such as root mean square error and Pearson correlation are inappropriate for evaluating the speaker-independence of models or the usefulness of combining datasets. We present a new evaluation for articulatory reconstruction which is independent of the articulatory data set used for training: the phone discrimination ABX task. We use the ABX measure to evaluate a Bi-LSTM based model trained on 3 datasets (14 speakers), and show that it gives information complementary to the standard measures, and enables us to evaluate the effects of dataset merging, as well as the speaker independence of the model.

Ewan Dunbar, Robin Algayres, Julien Karadayi et al.

We present the Zero Resource Speech Challenge 2019, which proposes to build a speech synthesizer without any text or phonetic labels: hence, TTS without T (text-to-speech without text). We provide raw audio for a target voice in an unknown language (the Voice dataset), but no alignment, text or labels. Participants must discover subword units in an unsupervised way (using the Unit Discovery dataset) and align them to the voice recordings in a way that works best for the purpose of synthesizing novel utterances from novel speakers, similar to the target speaker's voice. We describe the metrics used for evaluation, a baseline system consisting of unsupervised subword unit discovery plus a standard TTS system, and a topline TTS using gold phoneme transcriptions. We present an overview of the 19 submitted systems from 10 teams and discuss the main results.

R. Thomas McCoy, Tal Linzen, Ewan Dunbar et al.

Recurrent neural networks (RNNs) can learn continuous vector representations of symbolic structures such as sequences and sentences; these representations often exhibit linear regularities (analogies). Such regularities motivate our hypothesis that RNNs that show such regularities implicitly compile symbolic structures into tensor product representations (TPRs; Smolensky, 1990), which additively combine tensor products of vectors representing roles (e.g., sequence positions) and vectors representing fillers (e.g., particular words). To test this hypothesis, we introduce Tensor Product Decomposition Networks (TPDNs), which use TPRs to approximate existing vector representations. We demonstrate using synthetic data that TPDNs can successfully approximate linear and tree-based RNN autoencoder representations, suggesting that these representations exhibit interpretable compositional structure; we explore the settings that lead RNNs to induce such structure-sensitive representations. By contrast, further TPDN experiments show that the representations of four models trained to encode naturally-occurring sentences can be largely approximated with a bag of words, with only marginal improvements from more sophisticated structures. We conclude that TPDNs provide a powerful method for interpreting vector representations, and that standard RNNs can induce compositional sequence representations that are remarkably well approximated by TPRs; at the same time, existing training tasks for sentence representation learning may not be sufficient for inducing robust structural representations.

Ewan Dunbar, Xuan Nga Cao, Juan Benjumea et al.

We describe a new challenge aimed at discovering subword and word units from raw speech. This challenge is the followup to the Zero Resource Speech Challenge 2015. It aims at constructing systems that generalize across languages and adapt to new speakers. The design features and evaluation metrics of the challenge are presented and the results of seventeen models are discussed.

Rahma Chaabouni, Ewan Dunbar, Neil Zeghidour et al.

Recent works have explored deep architectures for learning multimodal speech representation (e.g. audio and images, articulation and audio) in a supervised way. Here we investigate the role of combining different speech modalities, i.e. audio and visual information representing the lips movements, in a weakly supervised way using Siamese networks and lexical same-different side information. In particular, we ask whether one modality can benefit from the other to provide a richer representation for phone recognition in a weakly supervised setting. We introduce mono-task and multi-task methods for merging speech and visual modalities for phone recognition. The mono-task learning consists in applying a Siamese network on the concatenation of the two modalities, while the multi-task learning receives several different combinations of modalities at train time. We show that multi-task learning enhances discriminability for visual and multimodal inputs while minimally impacting auditory inputs. Furthermore, we present a qualitative analysis of the obtained phone embeddings, and show that cross-modal visual input can improve the discriminability of phonological features which are visually discernable (rounding, open/close, labial place of articulation), resulting in representations that are closer to abstract linguistic features than those based on audio only.