skip to main content

Title: Self-supervised Semantic-driven Phoneme Discovery for Zero-resource Speech Recognition
Phonemes are defined by their relationship to words: changing a phoneme changes the word. Learning a phoneme inventory with little supervision has been a longstanding challenge with important applications to under-resourced speech technology. In this paper, we bridge the gap between the linguistic and statistical definition of phonemes and propose a novel neural discrete representation learning model for self-supervised learning of phoneme inventory with raw speech and word labels. Under mild assumptions, we prove that the phoneme inventory learned by our approach converges to the true one with an exponentially low error rate. Moreover, in experiments on TIMIT and Mboshi benchmarks, our approach consistently learns a better phoneme-level representation and achieves a lower error rate in a zero-resource phoneme recognition task than previous state-of-the-art self-supervised representation learning algorithms.
; ; ;
Muresan, Smaranda; Nakov, Preslav; Villavicencio, Aline
Award ID(s):
Publication Date:
Journal Name:
Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
Sponsoring Org:
National Science Foundation
More Like this
  1. Current leading mispronunciation detection and diagnosis (MDD) systems achieve promising performance via end-to-end phoneme recognition. One challenge of such end-to-end solutions is the scarcity of human-annotated phonemes on natural L2 speech. In this work, we leverage unlabeled L2 speech via a pseudo-labeling (PL) procedure and extend the fine-tuning approach based on pre-trained self-supervised learning (SSL) models. Specifically, we use Wav2vec 2.0 as our SSL model, and fine-tune it using original labeled L2 speech samples plus the created pseudo-labeled L2 speech samples. Our pseudo labels are dynamic and are produced by an ensemble of the online model on-the-fly, which ensures that our model is robust to pseudo label noise. We show that fine-tuning with pseudo labels achieves a 5.35% phoneme error rate reduction and 2.48% MDD F1 score improvement over a labeled-samples-only finetuning baseline. The proposed PL method is also shown to outperform conventional offline PL methods. Compared to the state-of-the-art MDD systems, our MDD solution produces a more accurate and consistent phonetic error diagnosis. In addition, we conduct an open test on a separate UTD-4Accents dataset, where our system recognition outputs show a strong correlation with human perception, based on accentedness and intelligibility.
  2. Human speech perception involves transforming a countinuous acoustic signal into discrete linguistically meaningful units (phonemes) while simultaneously causing a listener to activate words that are similar to the spoken utterance and to each other. The Neighborhood Activation Model posits that phonological neighbors (two forms [words] that differ by one phoneme) compete significantly for recognition as a spoken word is heard. This definition of phonological similarity can be extended to an entire corpus of forms to produce a phonological neighbor network (PNN). We study PNNs for five languages: English, Spanish, French, Dutch, and German. Consistent with previous work, we find that the PNNs share a consistent set of topological features. Using an approach that generates random lexicons with increasing levels of phonological realism, we show that even random forms with minimal relationship to any real language, combined with only the empirical distribution of language-specific phonological form lengths, are sufficient to produce the topological properties observed in the real language PNNs. The resulting pseudo-PNNs are insensitive to the level of lingustic realism in the random lexicons but quite sensitive to the shape of the form length distribution. We therefore conclude that “universal” features seen across multiple languages are really string universals, notmore »language universals, and arise primarily due to limitations in the kinds of networks generated by the one-step neighbor definition. Taken together, our results indicate that caution is warranted when linking the dynamics of human spoken word recognition to the topological properties of PNNs, and that the investigation of alternative similarity metrics for phonological forms should be a priority.« less
  3. Recently proposed self-supervised learning approaches have been successful for pre-training speech representation models. The utility of these learned representations has been observed empirically, but not much has been studied about the type or extent of information encoded in the pre-trained representations themselves. Developing such insights can help understand the capabilities and limits of these models and enable the research community to more efficiently develop their usage for downstream applications. In this work, we begin to fill this gap by examining one recent and successful pre-trained model (wav2vec 2.0), via its intermediate representation vectors, using a suite of analysis tools. We use the metrics of canonical correlation, mutual information, and performance on simple downstream tasks with non-parametric probes, in order to (i) query for acoustic and linguistic information content, (ii) characterize the evolution of information across model layers, and (iii) understand how fine-tuning the model for automatic speech recognition (ASR) affects these observations. Our findings motivate modifying the fine-tuning protocol for ASR, which produces improved word error rates in a low-resource setting.
  4. Abstract

    A longstanding debate has surrounded the role of the motor system in speech perception, but progress in this area has been limited by tasks that only examine isolated syllables and conflate decision-making with perception. Using an adaptive task that temporally isolates perception from decision-making, we examined an EEG signature of motor activity (sensorimotor μ/beta suppression) during the perception of auditory phonemes, auditory words, audiovisual words, and environmental sounds while holding difficulty constant at two levels (Easy/Hard). Results revealed left-lateralized sensorimotor μ/beta suppression that was related to perception of speech but not environmental sounds. Audiovisual word and phoneme stimuli showed enhanced left sensorimotor μ/beta suppression for correct relative to incorrect trials, while auditory word stimuli showed enhanced suppression for incorrect trials. Our results demonstrate that motor involvement in perception is left-lateralized, is specific to speech stimuli, and it not simply the result of domain-general processes. These results provide evidence for an interactive network for speech perception in which dorsal stream motor areas are dynamically engaged during the perception of speech depending on the characteristics of the speech signal. Crucially, this motor engagement has different effects on the perceptual outcome depending on the lexicality and modality of the speech stimulus.

  5. Abstract
    Speech processing is highly incremental. It is widely accepted that human listeners continuously use the linguistic context to anticipate upcoming concepts, words, and phonemes. However, previous evidence supports two seemingly contradictory models of how a predictive context is integrated with the bottom-up sensory input: Classic psycholinguistic paradigms suggest a two-stage process, in which acoustic input initially leads to local, context-independent representations, which are then quickly integrated with contextual constraints. This contrasts with the view that the brain constructs a single coherent, unified interpretation of the input, which fully integrates available information across representational hierarchies, and thus uses contextual constraints to modulate even the earliest sensory representations. To distinguish these hypotheses, we tested magnetoencephalography responses to continuous narrative speech for signatures of local and unified predictive models. Results provide evidence that listeners employ both types of models in parallel. Two local context models uniquely predict some part of early neural responses, one based on sublexical phoneme sequences, and one based on the phonemes in the current word alone; at the same time, even early responses to phonemes also reflect a unified model that incorporates sentence-level constraints to predict upcoming phonemes. Neural source localization places the anatomical origins of the differentMore>>