A Study on Extraction of Robust Parameters of Individuality for Speaker Recognition Based on Acoustic Theory of Speech Production

为了阐明声道解剖结构、声学参数和听觉感知之间的关系，我们开发了一种利用持续元音的磁共振图像测量三维声道形状的新方法。通过磁共振成像测量了三个成年男性、三个女性和一个男孩发出的五个日语元音的声道形状尺寸。系统地研究了男性、女性和儿童声道形状和声学参数的差异。结果表明，声道形状尺寸的不均匀性对声学特性没有本质影响。在这项研究中，从男性和女性的磁共振图像测量元音声道形状的个性进行了讨论。研究了受试者声道尺寸的差异及其对声学特性的影响。元音质量的感知相似性测试表明，在很大程度上依靠声道长度可以使女性和男性的元音更加标准化。在发音和声学水平上仔细比较了雄性的元音。结果表明，对于一个相同的男性元音，发音的“不变性”可能是声学参数（前三个形成峰频率F1， F2和F3，这在听觉感知中很重要），而不是发音模拟。较高的共振峰频率（F4, F5）是说话人个性的稳定因子。为了从语音信号中稳定地估计说话人的个体参数，引入了一种源-滤波器模型来表示语音产生，该模型将语音信号视为受声源激励的滤波器（声道）的输出。提出了一种新的语音分析方法，利用直接基于子空间的状态空间系统识别算法对模型参数进行分析。实验结果表明，该方法不仅能很好地估计声道高频峰参数，而且能很好地估计声源参数。除了说话人识别之外，本项目的成果可以应用于几乎所有的语音研究领域，如合成、感知、语音转换、编码、识别。少

In order to elucidate the relation among the anatomical structure of the vocal tract, the acoustic parameters, and the auditory perception, we have developed a new method to measure three dimensional vocal tract shapes from magnetic resonance (MR) images of sustained vowels. The dimensions of the vocal tract shapes of five Japanese vowels phonated by three adult males, three females and a boy are measured from the MR images. Differences of vocal tract shapes and acoustic parameters among male, female and child are systematically investigated. The result showed that the non-uniformity in the dimensions of the vocal tract shape had no essential effect on the acoustic characteristics.In this research individualities of the vocal tract shape of vowels measured from MR images of males and females were discussed. Differences in dimensions of the vocal tract of the subjects and their effects on acoustic characteristics were investigated. Perceptual similarity tests of vowel quality showed tha … More t normalization of vowels from females to males could be made by relying largely on the vocal tract length.Vowels of the males were carefully compared at the articulatory and acoustic levels. The result suggested that, for an identical vowel of males, the "invariance" of the phonation may be acoustic parameters (the first three formant frequencies F1, F2 and F3, which are important in auditory perception) rather than the articulatory simulation. It was also showed that the higher formant frequencies (F4, F5) are stable factors of speaker individuality.To stably estimate speaker individual parameters from a speech signal, a source-filter model was introduced to represent the speech production, in which a speech signal is regarded as the output of a filter (the vocal tract) excited by a sound source.A novel speech analysis method was proposed to analyze the model parameters by using a direct subspace-based state-space system identification algorithm. Experimental results showed that not only the vocal tract parameters including the higher formant frequencies but also the source parameters can be estimated quite well.In addition to speaker recognition, the results of this project can be expected to be applied in almost all the speech research areas, such as synthesis, perception, voice conversion, coding, recognition. Less