一种新的基于信息熵的带噪语音端点检测方法

A Novel Approach to Entropy-based Endpoint Detection of Noisy Speech

在自动语音识别和变速率语音编码技术中,语音端点检测是前端处理的一个重要环节。而在实际的噪声环境下,一些传统的端点检测方法已不适用。该文提出了一种新的基于信息熵的语音端点检测方法,该方法通过对语音信号的短时功率谱进行谱分析,由此构造熵函数作为端点检测的特征参数。实验结果表明,该方法在噪声环境下性能优于传统的基于能量的端点检测方法。而且相对于基于频谱谱熵的算法[1],在低信噪比(SNR<0dB)情况下,该文方法有更好的鲁棒性,可使平均检测精确度进一步提高约5%。

In the technology of speech recognition and variable bit rate speech coding, accurate determination of speech is a crucial part. Some traditional endpoint detection methods are ineffective in real noisy environments. This paper presents a novel entropy - based approach to speech endpoint detection. In the proposed method, we analyze the short - term power spectrum of speech signal, from which the entropy is derived and used as a feature in endpoint detection. Experimental results show that this method outperforms the traditional energy - based methods. Compared with the spectral entropy - based algorithm it possesses a better robustness in low SNR environments （ SNR 〈 0dB） and can improve the precision of endpoint detection about 5%.