采用子带长时信号变化特征的稳健语音活动检测

为提高语音活动检测(VAD)在低信噪比下的准确率,提出了一种基于子带长时信号变化特征的VAD算法.将语音信号转换到频域,并分解为几个不重复的子频带,对这些子带信号分别提取长时信号变化特征,然后采用GMM在线建立语音和非语音模型,以模型的似然比进行VAD判决.实验结果表明,算法在较低的信噪比下能够显著地提高语音活动检测的准确率,且在多种噪声环境和信噪比条件下具有较好的稳健性.应用于语音识别系统的实验表明,该算法能有效提高噪声环境下的语音识别率.

Robust Voice Activity Detection Using Sub-band Long-term Signal Variability Features

To improve the accuracy of voice activity detection (VAD) in low signal-to-noise ratio (SNR) environments, an algorithm based on the sub-band long-term signal variability features is proposed. Each frame of speech signal is first divided into several non-repetitive sub-bands and the long-term signal variability features of each sub-band are then calculated. The speech and non-speech models on an utterance-by-utterance basis are then trained on-line using GMM. A likelihood ratio detector is used for VAD decision at the final step. Experimental results show that the algorithm significantly increases the accuracy of voice activity detection at low SNR and provides better noise robustness in various noisy environments at any SNR. In addition, the speech recognition system utilizing the proposed VAD algorithm can achieve better performance in noisy environments.