Word Recognition using A Two-Dimensional Mel-cepstrum under Noisy Environments.

本研究的目的是为噪声环境下的文字识别提供一种新的方法。在这项研究中，白色噪音产生的计算机模拟和彩色噪音记录在名古屋站。提出了一种基于二维梅尔倒谱（TDMC）的日语十位数字非特定人识别方法。TDMC定义为在频域和时域上对Mel频率标度的对数谱进行二维傅里叶变换，由二维Mel对数谱的平均特征和动态特征组成.采用TMDC的语音分析合成系统及其估计提出了采用TMDC的语音分析合成系统的结构，以研究合成高质量语音所需的TDMC的大小。结果表明，所需的TDMC面积的频率小于约10 Hz。参考模式的输入语音的信噪比（SNR）的变化鲁棒性;在这项研究中，一组TDMC的噪声添加的参考模式与期望的SNR用于在噪声环境下的单词识别。实验结果表明，使用该参考模式集的识别方法比通常的方法更有效.一种对输入语音信噪比变化具有鲁棒性的词识别方法的距离测度提出了一种结合TDMC的动态和平均特征的距离测度。结果表明，在噪声环境下，动态特征比平均特征对单词识别更重要。

The purpose of this research is to offer a new method for word recognition under noisy environments. In this study white noise generated by computer simulation and colored noise recorded in the Nagoya station are used. A speaker- independent word recognition method of ten Japanese digits using a two- dimensional mel-cepstrum(TDMC) is proposed. TDMC is defined as the two- dimensional Fourier transform of mel-frequency scaled logarithm spectra in the frequency and time domains and consists of average features and dynamic features of the two-dimensional mel-log spectra, Experimental results in this study are shown as follows.1. Speech analysis-synthesis system using a TMDC and its estimation; The structure of speech analysis-synthesis system using a TMDC is proposed in order to study the size of the TDMC for synthesizing good quality speech. It is shown that the frequency of the required area of the TDMC is less than about 10Hz.2. Reference patterns robust for the variation of signal-to-noise ratio (SNR) of input speech; In this study a single set of TDMCs of noise-added reference patterns with desired SNR is used for word recognition under noisy environments. Experimental results show that a recognition method using this reference pattern set is more effective than a usual method.3. Distance measures for a word recognition method robust for the variation of SNR of input speech; Distance measures using a combination of dynamic and average features of the TDMC is proposed. It is shown that dynamic features are more important than average features for word recognition under noisy environment.