Recent advances in self-supervised learning have allowed automatic speech recognition (ASR) systems to achieve state-of-the-art (SOTA) word error rates (WER) while requiring only a fraction of the labeled data needed by its predecessors. Notwithstanding, while such models achieve SOTA results in matched train/test scenarios, their performance degrades substantially when tested in unseen conditions. To overcome this problem, strategies such as data augmentation and/or domain adaptation have been explored. Available models, however, are still too large to be considered for edge speech applications on resource-constrained devices; thus, model compression tools, such as knowledge distillation, are needed. In this paper, we propose three innovations on top of the existing DistilHuBERT distillation recipe: optimize the prediction heads, employ a targeted data augmentation method for different environmental scenarios, and employ a real-time environment estimator to choose between compressed models for inference. Experiments with the LibriSpeech dataset, corrupted with varying noise types and reverberation levels, show the proposed method outperforming several benchmark methods, both original and compressed, by as much as 48.4% and 89.2% in the word error reduction rate in extremely noisy and reverberant conditions, respectively, while reducing by 50% the number of parameters. Thus, the proposed method is well suited for resource-constrained edge speech recognition applications.
Self-supervised speech representation learning aims to extract meaningful factors from the speech signal that can later be used across different downstream tasks, such as speech and/or emotion recognition. Existing models, such as HuBERT, however, can be fairly large thus may not be suitable for edge speech applications. Moreover, realistic applications typically involve speech corrupted by noise and room reverberation, hence models need to provide representations that are robust to such environmental factors. In this study, we build on the so-called DistilHuBERT model, which distils HuBERT to a fraction of its original size, with three modifications, namely: (i) augment the training data with noise and reverberation, while the student model needs to distill the clean representations from the teacher model; (ii) introduce a curriculum learning approach where increasing levels of noise are introduced as the model trains, thus helping with convergence and with the creation of more robust representations; and (iii) introduce a multi-task learning approach where the model also reconstructs the clean waveform jointly with the distillation task, thus also acting as an enhancement step to ensure additional environment robustness to the representation. Experiments on three SUPERB tasks show the advantages of the proposed method not only relative to the original DistilHuBERT, but also to the original HuBERT, thus showing the advantages of the proposed method for ``in the wild'' edge speech applications.