SEASR: Speech Enhancement for Automatic Speech Recognition Systems using Convolution Recurrent Neural Network with Residual Connections

Automatic Speech Recognition (ASR) systems have become increasingly popular in recent years with applications like virtual assistants and auto-captions. However, one of the major hurdles is the noise in speech signals which significantly degrades ASR accuracy. A simpler solution is to enhance the speech signal using a noise reduction algorithm prior to ASR. The existing speech enhancement algorithms suffer from a trade-off between noise reduction and speech distortion, and the difficulty in estimating non-stationary noises. In this work, we present SEASR, a lightweight fully causal realtime speech enhancement network for improving ASR accuracy in noisy environments. Our network seamlessly removes non-stationary noises from speech at low SNRs demonstrating a relative improvement of 20-60% of accuracy across various ASR systems. We also showcase the performance and quality improvements of our model compared to other popular noise suppression methods.