论文部分内容阅读
为提高语音端点检测系统在低信噪(0 dB以下)下检测的准确率,提出了一种基于谱熵的端点检测算法。将每帧信号分为16个子带,选取频谱分布在250~3.5 kH z并且能量不超过该帧总能量90%的子带,计算经过语音增强后的子带能量以及各子带信噪比,根据各子带信噪比的不同调整其在整个谱熵计算过程中的权重,然后平滑谱熵,以最终的谱熵作为端点检测的依据。实验结果表明,此方法在较低的信噪比下能够显著地提高端点检测的准确率。对坦克噪声,检测效果明显优于G.729中的端点检测算法,即使在-5 dB的信噪比下,仍然可以达到95%以上的检测率。
In order to improve the detection accuracy of speech endpoint detection system under low signal noise (below 0 dB), an endpoint detection algorithm based on spectral entropy is proposed. The signal of each frame is divided into 16 subbands, the subbands whose spectrum is distributed at 250-3.5 kH z and whose energy does not exceed 90% of the total energy of the frame are selected, the subbands after speech enhancement are calculated, and the signal to noise ratio of each subband is calculated, According to the different signal-to-noise ratio of each sub-band, the weights of the whole spectral entropy are adjusted, and then the spectral entropy is smoothed. The final spectral entropy is taken as the basis of endpoint detection. Experimental results show that this method can significantly improve the accuracy of endpoint detection at a low SNR. For tank noise, the detection effect is obviously better than the endpoint detection algorithm in G.729, which can achieve the detection rate of more than 95% even at -5 dB signal-to-noise ratio.