Improving Noise Robustness of Automatic Speech Recognition via Parallel Data and Teacher-student Learning

For real-world speech recognition applications, noise robustnessis still a challenge. In this work, we adopt the teacherstudent(T/S) learning technique using a parallel clean andnoisy corpus for improving automatic speech recognition(ASR) performance under multimedia noise. On top of that,we apply a logits selection method which only preserves the khighest values to prevent wrong emphasis of knowledge fromthe teacher and to reduce bandwidth needed for transferringdata. We incorporate up to 8000 hours of untranscribed datafor training and present our results on sequence trained modelsapart from cross entropy trained ones. The best sequencetrained student model yields relative word error rate (WER)reductions of approximately 10.1%, 28.7% and 19.6% on ourclean, simulated noisy and real test sets respectively comparingto a sequence trained teacher.