基于小波分析的梅尔频率倒谱参数?

采用HMM模型和MFCC参数的语音识别器对普通话中声母音素的区分度不够理想,而在基于识别的计算机辅助发音教学系统中的辅音的识别具有特别重要的意义.考虑到相同发音位置不同发音方式的声母音素变化较快且高频信息较多,本文将小波分析的方法引入到提取梅尔频率倒谱参数(MFCC)的过程当中,来提高信号高频部分的时域分辨率,提出了基于小波分析的梅尔倒谱参数MFCC_Wavelet.结合高低频不同分帧方式的MFCC_Wavelet参数与HMM模型的语音识别器,本文测试了MFCC和MFCC_Wavelet两种参数在4类发音中的区分性,实验结果表明,在相同发音位置不同发音方式、塞音与不塞音、送气音与不送气音及擦音与不擦音4类发音错误中,MFCC_Wavelet的总体效果好于MFCC.

Wavelet analysis based Mel frequency cepstrum parameters

Changing rapidly over time and with higher frequency,most consonants in Chinese Mandarin need shorter analysis frame length in automatic speech recognition (ASR).In contrast,longer frame suits vowels which are comparatively stable and with lower frequency distribution.A new speech feature MFCC-Wavelet is introduced here combining wavelet analysis with Mel frequency cepstrum coefficient (MFCC).It has higher time resolution in high frequency like wavelet analysis,and possesses Mel frequency resolution of MFCC satisfying both requirements of consonant and vowel recognition.Experiments showed better performance than MFCC to differentiate plosive/non-plosive,fricative/non-fricative and aspirated/non-aspirated phonemes in Chinese Mandarin recognition.These are important specifically in ASR-based computer-assisted pronunciation teaching (CAPT).