A Dual Target Neural Network Method for Speech Enhancement

Changhuai You, Minghui Dong, Lei Wang

Abstract: In this paper, we introduce a neural network based speech enhancement method for heavy noisy scenarios. Comprising bidirectional long-short term memory (BLSTM), convolutional neural network (CNN) and deep feedforward neural network (DNN), a neural network architecture with a novel dual target scheme is proposed. The dual target scheme leverages the speech features from both linear and Mel spectral domains in order to retrieve very weak speech spectrum hidden in heavy noise. Experiments show that our proposed neural network architecture outperforms many traditional enhancement methods in terms of objective distortion measure and quality measure.

External IDs:dblp:conf/ialp/YouDW23