Linear prediction-based dereverberation with very deep convolutional neural networks for reverberant speech recognition

Abstract

Convolutional neural networks (CNNs) have been shown to improve classification tasks such as automatic speech recognition (ASR). Furthermore, the CNN with very deep architecture lowered the word error rate (WER) in reverberant and noisy environments. However, DNN-based ASR systems still perform poorly in unseen reverberant conditions. In this paper, we use the weighted prediction error (WPE)-based preprocessing for dereverberation. In our experiments on the ASR task of the REVERB Challenge 2014, the WPE-based processing with eight channels reduced the WER by 20% for the real-condition data using CNN acoustic models with 10 layers.