基于LPC美尔倒谱特征的带噪语音端点检测

Endpoint Detection of Noisy Speech Based on LPC Cepstrum

复杂的噪声环境是语音识别系统在实际应用中性能下降的原因之一,识别预处理中的带噪端点检测作为关键技术,其性能的优劣某种程度上决定了识别率的高低。笔者提出了基于LPC美尔倒谱特征的带噪端点检测方法,对语音信号分高低频段分别提取LPC美尔倒谱特征分析,根据Mel倒谱距离判决,采用自适应噪声估计,实验结果表明,该方法计算效率较高,低信噪比下有较好的检测性能。

One of the causes reducing the capacity of speech recognition(SR)systems in practical use is the complex noisy environment. As an important technology in the preparation of SR, the endpoint detection of noisy speech's accuracy determines the SR rates in some degree. In this article, a method based on LPCCMCC for endpoint detection of noisy speech signal is proposed. It filtered the speech signal into high and low frequency bands, then analyzed LPC Mel cepstrum feature  LPCCMCC  respectively, and decided the endpoint by Mel cepstral distance, making the estimation of noise adaptive. The experiments show that higher efficiency and good detection capability can be obtained under the condition of low SNR.