Abstract: Aiming at the question that the performance of speech segmentation declines distinctly in noise environment，this paper proposes a new speech segmentation method for continuous speech signal.The method doesn’temploy a single method for endpoint detection，but combines several different results derived from different endpoint detection methods based on fractal dimension，cepstral feature and HMM model，using a candidate selection approach to get the final boundary in order to segment the continuous speech signal.The experimental results show that the proposed approach rather improves the speech segmentation accuracy.

Key words: speech segmentation, cepstral feature, fractal dimension, Hidden Markov Model（HMM）, candidate selection, noise environment

摘要： 针对在噪声背景下连续语音信号的语音分割性能会明显下降的问题，提出了一种针对连续语音信号分割的新方法。该方法不再采用单一的端点检测方法，而是将基于分形维数的端点检测方法，基于倒谱特征的端点检测方法，基于HMM的端点检测方法等多种不同方法下得到的端点检测结果，通过投票选择的方式，得到最终的端点检测结果，从而达到对连续语音信号进行分割的目的。实验结果表明，该方法较明显地提高了语音分割的准确性。

关键词: 语音分割, 倒谱特征, 分形维数, 隐马尔科夫模型（HMM）, 投票选择, 背景噪声