Classification of Japanese Consonants in terms of Structured Modeling and Geometrical Interpretation of Dynamic Characteristics in Transient Parts

近些年来，语音识别中的瞬时词性信号一直是语音识别领域的研究热点。Furui提出了增量倒谱系数来提取语音信号的动态特征，取得了比以往更高的识别率[1986]。然而，应当注意，包括动态特性的传统方法依赖于以语音信号的格式为特征的线性预测。我们使用自然观测系统的结构模型进行语音建模，并利用几何解释对暂态部分进行分类。自然观测滤波器(NOF)由一系列具有相同截止频率的一阶滤波器组成：第一级的低通滤波器，然后是高通滤波器。NOF可以利用一阶滤波器的每一级的输出的线性组合，采用适当的系数来重构输入信号。我们称它们为自然观测系数(NOC)。我们可以根据具有多个截止频率的NoC的时间历程(NoC模式)来分析信号，因为每个NoC都是截止频率的函数。我们应用日语辅音、摩擦和停顿的NOC模式来研究语音信号的暂态部分。因此，我们得到了可区分的NOC模式，这些模式可能反映了日语辅音、摩擦和停顿以及紧随其后的元音的瞬时结构。另一方面，我们使用了对过渡部分的几何解释。确定了p维Hilbert空间中的二维平面，使得系数向量的轨迹尽可能地变得笔直。我们使用了三种类型的参数，包括线性预测系数、LPC倒谱系数和NoCs，它们都是结构化参数。因此，与其他两家相比，国家石油公司表现出了一条相当直线的曲线。限制在LPC倒谱系数范围内，从元音到下一个元音的平稳渐进轨迹有助于提高识别率。

Transient part of speech signals has been of great interest these years in speech recognition. Furui proposed delta cepstrum coefficients to extract a dynamic characteristic of speech signals and achieved higher recognition rate than ever [1986]. It should be noted, however, that conventional approaches including the dynamic characteristics depend on a linear prediction that features formats of speech signals. We employed a structural model named the natural observation system for speech modeling and a geometrical interpretation to classify the types of transient parts.A natural observation filter (NOF) consists of a series of first order filters with the same cutoff frequency : a low pass filter of a first stage followed by high pass filters. The NOF can reconstruct an input signal with a linear combination of outputs at each stage of first order filters, employing appropriate coefficients. We called them natural observation coefficients (NOCs). We can analyze signals by the time courses of NOCs with multi-cutoff frequencies (NOC pattern), because each NOC is a function of the cutoff frequency. We applied the NOC pattern for Japanese consonants, fricatives and stops, to study the transient parts of speech signals. Consequently, we obtained distinguishable NOC patterns that probably reflect an instantaneous structure of Japanese consonants, fricatives and stops, and following vowels.On the other hand, we used a geometrical interpretation of the transient parts. A two-dimensional plane in the p-dimensional Hilbert space was determined so that the trajectory of coefficients vector becomes as straight as possible. We use three types of parameters, including linear prediction coefficients, LPC cepstrum coefficients, and NOCs which are structured parameters. As a result, NOCs show a rather straight line than the other two. Limiting to LPC cepstrum coefficients, the smoothly progressive trajectory to a following vowel contributed to the high recognition rate.