关键词: 声谱图, 小波矩, 特征提取, 向后传播（BP）网络, 声音识别

Abstract: Currently, most of voice recognition systems can reach a very high recognition rate in noiseless conditions. However, the sound recognition sharply declines in noisy environment. Given this problem, this paper puts forward a voice recognition method based on wavelet moment and BP network. The system generates sonograms according to sound signals. Then, it uses wavelet moment to extract features from spectrogram, and select the representative characteristic parameters. It uses BP neutral network to classify and recognize sounds based on characteristic parameters. The experimental results show that the method can achieve good results in noise environment with different types of noise or different SNR, which overcomes the defect of the low recognition rate under low SNR.

Key words: sonograms, wavelet moment, feature extraction, Back Propagation（BP） neural network, voice recognition