检测语音端点及基音的概率模型及方法  

Endpoint detection and pitch determination method based on a probability model

作　　者：胡波[1] 肖熙[1] 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《清华大学学报（自然科学版）》2013年第6期749-752,共4页Journal of Tsinghua University(Science and Technology)

摘　　要：在传统的语音识别系统中,语音端点检测和基音提取是2个分开的步骤。将2个步骤合二为一将有助于提高误别准确性、降低系统复杂度。该文使用了一种语音时域分析方法,它使用概率模型描述语音幅度分布规律,并使用隐Markov模型(hidden Markov model,HMM)描述语音中的状态转换。使用新方法可以同时完成对语音端点的检测、清浊音的判断与基音频率计算。实验表明:这种算法在10dB以上信噪比的条件下可以得到准确的基音频率和端点位置。Endpoint detection and pitch extraction are separated steps in traditional speech recognition systems.Combining these two steps together will improve precision and reduce complexity.A time domain analysis method is developed to describe the speech signal amplitude with a probability model and to model the state change of the speech with a hidden Markov model.The method simultaneously extracts the speech pitch and endpoints.Tests show that this algorithm can precisely detect the speech pitch frequency and endpoints for databases with a signal to noise ratio（SNR） more than 10 dB.

关 键 词：端点检测 基音提取 概率模型 统计方法 隐Markov模型(HMM) 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]