A Novel Approach to Speech Enhancement Based on Deep Neural Networks

Abstract

Minimum mean-square error (MMSE) approaches have been shown to achieve state-of-the-art performance on the task of speech enhancement. However, MMSE approaches lack the ability to accurately estimate non-stationary noise sources. In this paper, a long short-term memory fully convolutional network (LSTM-FCN) is utilized to accurately estimate a priori signal-to-noise ratio (SNR) since the speech enhancement performance of an MMSE approach improves with the accuracy of the used a priori SNR estimator. The proposed MMSE approach makes no assumptions about the characteristics of the noise or the speech. MMSE approaches that utilize the LSTM-FCN estimator are evaluated using the mean opinion score of the perceptual evaluation of speech quality (PESQ) and the short-time objective intelligibility (STOI) measures of speech. The experimental investigation shows that the speech enhancement performance of an MMSE approach that utilizes LSTM-FCN estimator significantly increases.

Keywords

• long short-term memory
• machine learning
• mean square error methods
• recurrent neural networks
• speech enhancement