一种基于频域特征和过渡段判决的端点检测算法

为了提高语音端点检测的准确性,增强端点检测算法在噪声环境下的鲁棒性,提出两种新的端点检测参数。其中,基于临界频带的谱熵参数综合考虑了人耳对语音的感知特性以及语音信号和噪声信号的频域分布差异,差值频域能量参数考虑了语音帧和无声帧在频域上的能量差异。结合两种参数的优点,构成一种鲁棒的端点检测参数,同时,为了避免因阀值判决的单一性而产生误判,在端点检测过程中加入了基于特征分布统计的过渡段判决。试验结果表明,本研究提出的语音端点检测算法对语音帧和无声帧具有较好的区分性,在不同噪声且信噪比较低情况下,端点检测准确率相比传统抗噪端点检测算法均有所提升,特别是在非平稳噪声下,准确率提升超过5%。

An endpoint detection algorithm based on frequency-domain characteristics and transition fragment judgment

In order to improve the accuracy of speech endpoint detection as well as enhance robustness of the endpoint detection algorithm in noisy environment, two new endpoint detection parameters were proposed. The spectrum entropy based on critical band took both perceptual characteristics of the human auditory system and the differences between speech and noise signals in frequency domain distribution into account, as well as the minus frequency-domain energy parameter paid attention to the difference between speech frames and silence frames in frequency energy. The advantages of those two parameters were combined to constitute a robust endpoint detection parameter. Meanwhile, in order to avoid the miscarriage of judgment caused by the unitary threshold, the transition fragment judgment based on statistics of characteristics distribution was applied. The experiment results showed that the endpoint detection algorithm had better discrimination for speech frames and silence frames, the algorithm could carry out better accuracy than other conventional anti-noisy endpoint detection algorithms under different and low signal-to-noise ratio noisy environments, especially in the case of non-stationary noise, the accuracy improved by more than 5%.