Analysis of Speech Enhancement Algorithm in Industrial Noise Environment

摘要

With the application of intelligent speech technology in the field of intelligent manufacturing, it is necessary to propose a speech enhancement algorithm that is suitable for industrial noise. In order to solve this problem, four types of optimized speech enhancement algorithms were selected based on the noise reduction capability and the degree of distortion: multi-band spectral subtraction algorithm, Wiener algorithm based on a priori SNR estimation, minimum mean square error of log-spectral amplitude estimation and subspace method based on Eigen-value decomposition (EVD) embedded pre-whitening. In low SNR conditions of -5 dB, 0 dB, and 5 dB, the speeches with industrial noise were used for experimental analysis. The experimental results were evaluated by the segmented SNR, perceptual evaluation of speech quality (PESQ), and time-domain waveforms, indicating that the Wiener algorithm based on a priori SNR estimation eliminates the noise better and improves speech quality higher. So it is more suitable for industrial noise environments than the other three algorithms.