基于人耳听觉特性的语音识别预处理研究  

Pretreatment of Speech Recognition Based on Human Auditory Characteristics

作　　者：张毅[1] 黎小松 罗元[1] 吴承军[1] 

机构地区：[1]重庆邮电大学信息无障碍工程研发中心,重庆400065

出　　处：《计算机仿真》2015年第12期322-326,共5页Computer Simulation

基　　金：国家科技部国际合作项目(2010DFA12160);重庆市科技攻关项目(CSTC:2010AA2055);重庆市科研项目(KJ13051)

摘　　要：在人耳听觉语音识别优化过程中,由于在噪声环境下传统语音识别预处理过程不能得到高信噪比的语音信号,使识别率下降。为此结合人耳听觉特性,提出一种基于人耳的听觉选择能力即"鸡尾酒会效应"的语音分离技术应用到语音识别预处理过程。含噪声的语音信号经过耳蜗基底膜模型进行频谱分析,再通过上橄榄核模型进行语音信息提取,最后在下丘脑细胞模型中完成语音分离。分离得到更纯净的语音后,对语音信号提取梅尔频率倒谱(MFCC)参数,并建立隐马尔可夫(HMM)声学模型来验证语音识别效果,实验结果表明:在噪声环境下,相比于传统抗噪方法,改进方法具有更好的抗噪效果,表明上述语音识别系统具有更好的鲁棒性。Combining with the human auditory characteristics, a speech separation technique was put forward based on the auditory ability of choice namely ＂cocktail party effect＂, and was applied to speech recognition pretreat- ment process. The spectral analysis carried out with noisy speech signal through the cochlea model. Then, voice in- formation was extracted by olive core model. Finally, the Speech Separation was completed through hypothalamus cell model. After getting more pure voice by separation, Mel - frequency cepstrum coefficients （MFCC） parameters of the voice signal were extracted, and Hidden Markov Model （HMM） was established to verify the effect of speech rec- ognition. The results indicate that under the noise environment, compared with the traditional method, this method has better effect of anti - noise ability and this speech recognition system has better robustness.

关 键 词：语音识别 人耳听觉特性 语音分离 预处理 

分 类 号：TP242.63[自动化与计算机技术—检测技术与自动化装置]