汉语数字耳语音识别研究

Speech Recognition of Chinese Whispered Speech

耳语音识别可应用于国家安全的某些特殊需要。运用双门限法对语音样本进行端点检测,通过实验分别找出短时能量、短时过零率的高低门限4个参数的最佳取值。深入分析研究参数的抗噪问题,在MFCC参数中引入短时能量、一阶差分、二阶差分等参数,增强MFCC的抗噪性。研究表明,在隐马尔可夫模型中,MFCC和LPCC联合运用讨论识别效果要远优于独立参数。

The whispered speech recognition even can be applied in the field of national security. In this paper,the characteristics of whispered speech in physiology and acoustics are introduced. The whispered speech is a noise sound source,the resonance peaks are offset,to recognize it more difficult than normal speech. The dual- threshold method of endpoint detection of voice samples is used,respectively,through experiments to identify the best value of the four parameters of short- time energy,short- time zero- crossing rate threshold. Depth analysis of the parameters of anti- noise problem; the introduction of short- time energy,first- order differential,second- order differential parameters and any other parameters in MFCC is made to enhance the anti- noise ability. The effect on recognition of joint use MFCC is much better than that and LPCC in HMM.