一种基于幅度谱偏度的语音激活检测算法

根据语音信号偏离高斯分布程度大而背景噪声信号偏离高斯分布程度小这一特征,提出一种改进的以语音短时幅度谱偏度为特征参数区分语音段和噪声段的语音激活检测算法,并应用到2.4 kbit/s混合激励线性预测(mixed excitation linear prediction,MELP)声码器中.通过与自适应多速率(adaptive multi-rate,AMR)语音编码标准中的语音激活检测算法相比较,该算法复杂度较小,且对背景噪声服从高斯分布的语音信号具有更好的端点检测性能.实现了可变速率MELP声码器的平均输出码率下降为1.9 kbit/s,通过非连续传输后合成的语音具有良好的舒适性和连续性.

A voice activity detection algorithm based on the spectral skewness

According to the characteristic that the speech signals deviate the Gaussian distribution larger than the background noise signals, this paper proposes an improved voice activity detection algorithm which uses the short-time spectral skewness as a characteristic parameter to distinguish the speech segments and non-speech segments, and then applies it to the 2.4kb/s mixed excitation linear prediction vocoder. Compared to the voice activity detection algorithm in AMR, the voice activity detection algorithm in this paper has a lower complexity and a better detection effect to the speeches mixed with Gaussian noise. The average coding rate of the MELP variable rate vocoder is reduced to 1.9 kbit/s, and the synthesized speeches applied in the discontinuous transmission have good comfort and continuity in the subjective hearing.