Study And Implementation Of Chinese Speech Feature Extraction And Analysis Tool

Acoustic feature extraction is one of key technologies of speech recognition and speaker recognition, which purpose is to transform the input speech signal into a reduced representation set of features through digital signal processing. Currently, the extraction of new features which reflect the characteristics of human auditory perception and are more noise-robust is a hot topic of speech recognition and speaker recognition. In recent years, world-wide researches on the speech feature extraction have been mostly focusing on English, rather than Chinese. Since comparing with English, Chinese has many different characteristics, it is necessary to intensify the research of Chinese speech feature extraction. In this paper, a great deal of research on Chinese speech signal analysis and feature extraction have been conducted, in detail, the following are included:1. Display method of spectrogram in Matlab platform is studied.2. Theory of auditory pitch perception and display method of Mel scale frequency spectrogram in Matlab platform is studied, and the corresponding algorithm is proposed.3. Theory of critical band is studied, and a filter group which contains 20 FIR filters is designed to simulate the 20 critical bands of human basilar membrane which cover the range of 200-9500Hz frequency.4. Short time energy and short time average zero crossing rate features which are often used in Chinese speech syllable segmentation are studied and the corresponding algorithms in Matlab is proposed.5. Evaluation methods of speaker features through statistics method are studied, and the corresponding algorithm in Matlab is proposed.6. A tool which integrates above functionalities are designed and implemented.7. A large number of speaker speech samples are collected, experiments are conducted to evaluate the performance of MFCC feature and LPC feature which are most commonly used in the Chinese speaker recognition systems, and to evaluate the performance of each dimension of MFCC feature. Speechlab is expected to be a handy tool on the research of Chinese speech signal analysis and feature extraction.