Noise-robust Attention Learning for End-to-End Speech Recognition

We propose a method for improving the noise robustness of an end-to-end automatic speech recognition (ASR) model using attention weights. Several studies have adopted a combination of recurrent neural networks and attention mechanisms to achieve direct speech-to-text translation. In the real-world environment, however, noisy conditions make it difficult for the attention mechanisms to estimate the accurate alignment between the input speech frames and output characters, leading to the degradation of the recognition performance of the end-to-end model. In this work, we propose noise-robust attention learning (NRAL) which explicitly tells the attention mechanism where to "listen at" in a sequence of noisy speech features. Specifically, we train the attention weights estimated from a noisy speech to approximate the weights estimated from a clean speech. The experimental results based on the CHiME-4 task indicate that the proposed NRAL approach effectively improves the noise robustness of the end-to-end ASR model.