Abstract: For speaker dependent speech recognition, noise can seriously affect to the extraction of features, thus leads to low speech recognition rate. To address the issue, this paper applies spectral subtraction for de-noising. Based on the visualized features of speech spectrogram, the pulse coupled neural network is used to extract the feature parameters for the recognition system. Experiment results show that the proposed method can preferably reduce the noises and achieve a significantly improved recognition rate for a speaker dependent recognition task.

Key words: speech recognition, pulse coupled neural network, spectral subtraction

摘要： 在特定人语音识别系统中，噪声严重影响语音特征提取，并导致语音识别率明显下降。针对在噪声环境下语音识别率偏低的问题，通过谱减法去除语音信号噪声，并根据语音信号语谱图可视化的特点，运用脉冲耦合神经网络从语音信号的语谱图中提取熵序列作为特征参数进行语音识别。实验结果表明，该方法能较好地去除语音信号中的噪声，并能使在噪声环境下的特定人语音识别系统具有较好的识别效果。

关键词: 语音识别, 脉冲耦合神经网络, 谱减法