A perceptual model of speech based on real-time speaker adaptation

感知实验表明，不同元音的说话者可以被正确识别，准确率在80%以上。对632位说话人发出的语音信号进行分析，提出了一种基于语音信号正弦表示的分析方法。此外，语音频谱的余弦展开及其系数的二次组合被证明是元音感知的有效特征。结果支持了听觉外围区两步突触结合提取元音感知特征的假设。

Perceptual experiments indicated that speakers of different vowels could be correctly identified with accuracy of more than 80%. Analyzing speech signals uttered by 632 speakers, a new analysis method was proposed on the basis of the sinusoidal representation of speech signal. Further, cosine expansion of speech spectra and the quadratic combination of their coefficients were shown to be effective features for vowel perception. The result supports the hypothesis that perceptual features for vowel might be extracted by two-step synaptic combination in auditory periphery.