基于子带双特征的自适应保留似然比鲁棒语音检测算法

Adaptively Reserved Likelihood Ratio-based Robust Voice Activity Detection with Sub-band Double Features

为了进一步提高低信噪比下语音激活检测（VAD）的准确率,该文提出一种基于子带双特征的自适应保留似然比鲁棒语音激活检测算法。算法采用子带归一化最大自相关函数与子带归一化平均过零率双重特征设置频率分量似然比的保留权值,同时利用已过去固定时长的VAD判决结果及对应的子带特征参数自适应地估计似然比的保留阈值。实验结果表明,此算法的VAD检测准确率相比原保留似然比算法在10 d B,0 d B和-10 d B平稳白噪声下分别提高了1.2%,7.2%和8.1%,在10 d B和0 d B非平稳Babble噪声下分别提高了1.6%和3.4%。当其被用于2.4 kbps低速率声码器系统时,合成语音的感知语音质量评价（PESQ）比原声码器系统在白噪声下提高了0.098~0.153,在Babble噪声下提高了0.157~0.186。

In order to improve the correct rate of Voice Activity Detection （VAD） in low Signal Noise Ratio （SNR） environment, the paper presents an adaptive reserved likelihood ratio VAD method, which is based on sub-band double features. The method employs sub-band auto correlate function and sub-band zero crossing rate in the process of setting reserved weight. Reserved threshold is estimated adaptively according to the passed VAD results and their sub-band feature parameters. The experiment shows its promising performance in comparison with similar algorithms, the VAD correct rate is improved by 1.2%, 7.2%, and 8.1% respectively in 10 dB, 0 dB, and -10 dB stationary white noisy environment, 1.6% and 3.4% respectively in 10 dB and 0 dB non-stationary Babble noisy environment. The method is also applied to 2.4 kbps low bit rate vocoder and the Perceptual Evaluation of Speech Quality （PESQ） is improved by 0.098-0.153 in white noisy environment, 0.157-0.186 in Babble noisy environment.