Abstract: To enhance the detection performance of statistical model-based Voice Activity Detection (VAD) using likelihood ratio test, an improved VAD was proposed by utilizing the correlation between tandem speech frames. First a priori Signal-to-Noise Ratio (SNR) was estimated using recursive estimation method based on the result of the previous speech frame instead of the traditional decision-directed method. Secondly double thresholds were designed by depending on the previous frame's detention result. Finally a detection rule was presented based on two-state Hidden Markov Model (HMM) coupled with double thresholds. The experimental results show that the inter-frame correlation based VAD scheme gets better performance than the Sohn's VAD.

Key words: voice activity detection, statistical model, correlation, likelihood ratio test, a priori SNR, threshold

摘要： 为了提高统计模型似然比测试的语音活动检测(VAD)的检测性能,利用前后语音帧间存在的统计相关特性,提出一种改进VAD算法。通过前帧语音频谱分量对先验信噪比进行递归估计,然后利用前一帧的语音检测状态来设计判决阈值,建立了双阈值隐马尔可夫模型语音活动判决规则。实验表明,此帧间相关性VAD算法的检测指标值优于Sohn算法。

关键词: 语音活动检测, 统计模型, 相关性, 似然比测试, 先验信噪比, 阈值