The Development Of Speech Signal Featrue Detection System

Feature extraction is an important part in signal processing, it has big significance for speaker recognition system and speech synthesis system. As each person has his or her own unique organ to produce speech, we need personalized features to distinguish different speakers. After fifty years’research, a lot of efficient features have been proposed in the area of speech processing. Such as, LPCC, MFCC, and PLP.One should firstly analyze the generative model if he wants to process speech signals. By deeply analysis of how to generate speech in human body, efficient generative model can be built. Then, more efficient parameters will be estimated for improving the similarity between manual synthetic and physical system of speech. In various kinds of synthetic systems, we always need to extract different features to divide which class that a section of speech belongs to. We can export three kinds of features, as in time domain, frequency domain, and cepstrum domain. Using such features, there will be much convenience in the future work. Besides, we should always have general understanding of the uniqueness of different phonemes. Then, the speech database can be built more effectively.In this paper, the generative model of speech has been researched, as well as multiple algorithms for extracting features. After research, some tests have been realized. Finally, a system which is based on Sphinx platform has been developed. It has the functions mainly for different features, voice activity detection, and pitch tracking. It can be a convenient tool in the research of speech signal processing.