Knowledge Distillation-Based Training of Speech Enhancement for Noise-Robust Automatic Speech Recognition

Abstract

This paper addresses the training issues associated with neural network-based automatic speech recognition (ASR) under noise conditions. In particular, conventional joint training approaches for a pipeline comprising speech enhancement (SE) and end-to-end ASR model surfer from a conflicting problem and a frame mismatched alignment problem because of different goals and different frame structures for ASR and SE. To mitigate such problems, a knowledge distillation (KD)-based training approach is proposed by interpreting the ASR and SE models in the pipeline as teacher and student models, respectively. In the proposed KD-based training approach, the ASR model is first trained using a training dataset, and then, acoustic tokens are generated via K-means clustering using the latent vectors of the ASR encoder. Thereafter, KD-based training of the SE model is performed using the generated acoustic tokens. The performance of the SE and ASR models is evaluated on two different databases, noisy LibriSpeech and CHiME-4, which correspond to simulated and real-world noise conditions, respectively. The experimental results show that the proposed KD-based training approach yields a lower character error rate (CER) and word error rate (WER) on the two datasets than conventional joint training approaches, including multicondition training. The results also show that the speech quality scores of the SE model trained using the proposed training approach are higher than those of SE models trained using conventional training approaches. Moreover, the noise reduction scores of the proposed training approach are higher than those of conventional joint training approaches but slightly lower than those of the standalone-SE training approach. Finally, an ablation study is conducted to examine the contribution of different combinations of loss functions in the proposed training approach to SE and ASR performance. The results show that the combination of all loss functions yields the lowest CER and WER and that tokenizer loss contributes more to SE and ASR performance improvement than ASR encoder loss.INDEX TERMS Noise-robust automatic speech recognition, speech enhancement, knowledge distillation, teacher-student model, acoustic tokenizer, K-means clustering. I. INTRODUCTIONRecent advancements in neural network architectures and training approaches have demonstrated consistent progress, enhancing capabilities in not only image and natural language processing but also audio and speech signal processing. Speech processing encompasses the comprehensive analysis, synthesis, and recognition of speech, including speaker verification, speech separation, speech enhancement (SE), speech synthesis, and automatic speech recognition (ASR) [1], [2]. ASR has gained considerable attention, particularly in voice-based information retrieval systems, chatbots, and automated transcription systems [3]. Moreover, the interest in ASR, specifically for real-world scenarios, such as closed This article has been accepted for publication in IEEE Access.