一种基于共振峰恢复和Ｍellin变换的非特定人语音特征提取方法

针对非特定人语音识别中的声道长度归一化问题，首先研究一种能够去牛基音激励的，基于自相关估计的共振峰（Ｆormart) 频谱恢复方法，说明了不同说话人发同一元音时的频谱互为尺度化的关系，以及它们与同一说话人发不同元音时频谱的差别，然后结合具有尺度不变性的Ｍellin变换，提出了一种适用于非特定人的语音特征提取方法，在实难中，对从非特定人收集的２０个汉语元音，分别提取了其ＦＦＴ倒谱，Ｍel倒谱，ＦＦＴ－Ｍellin倒谱及本文Ｆormant-Mellin倒谱，并用一种很直观的Ｆ－roato分辨率准则进行了性能评价，结果表明，无论是对纯净的，还是对带附加白噪声的发音样本，本文由共振峰恢复和Ｍellin变换相结合得到的语音特征都具有较高的分辨率。

A Novel SpeakerNormalization Method Based on Formant Recovery and Mellin Transform

To make the speech features insensitive to the vocal tract lengths among different speakers, the Mellin transform which is scale invariant and the formant recovery based on autocorrelation are used.Firstly, instead of direct FFT, a method based on the autocorrelation estimation is used to get the formant envelope. It suppresses the effects of pitch excitation, and is functionally equivalent to time-smoothed Wigner distribution. Then, the scale invariance and the discrete implement of the modified Mellin transform is introduced. Finally, a generalized F-ratio criterion is given as an objective rule to measure the separability between different vowel classes.In the experiment, 20 Chinese vowels spoken by different persons are used, and FFT cepstrum, FFT-Mel cepstrum, FFT-Mellin cepstrum and Formant-Mellin cepstrum are extracted for them respectively. F-ratio seperability is calculated for them with different vector dimensions. In addition, the same experiment is carried out for utterances with 6 dB artificial Gaussian white noise. Results show that formant-Mellin cepstrum not only has the highest separability rate, but also is more robust to noise than FFT-Mellin cepstrum.