【摘要】 目的研究基于涨落复杂性测度的语音特征提取,提高低信噪比语音端点检测的正确率和鲁棒性,从而改善语音处理和分析的性能。方法分析状态空间分割方法、窗长以及分区数对检测性能的影响。采用基于信息增益的复杂性行为度量,对含不同噪声类型,以及不同信噪比的各种中英文语音样本进行了对比实验。结果在低信噪比情况下,涨落复杂性测度比广泛应用的谱熵方法更有效。结论涨落复杂性测度技术可以较好地实现在动态噪声环境下对语音端点的检测。该方法鲁棒性好,算法实时性高。更多还原

【Abstract】 Objective To find a useful index for real-time detecting of speech endpoint and improving the performance of speech processing under low SNR by analyzing fluctuation complexity of speech signals. Method The influence of state space partition method, window size and partition numbers on detecting performance was analyzed. The comparison experiments of speech signals corresponding to different SNR and noise type was designed using the measure of complexity behaviors based on the information gain.Result It was found that fluctuation complexity was more effective in detecting low-SNR speech than spectral entropy. Conclusion Fluctuation complexity is a valid feature to make speech/non-speech decision for the low SNR cases. The presented method can achieve robust performance and has a good real-time behavior.更多还原