Speech Signal Prosodic Features Extraction And Its Application Research

The prosodic feature of speech signal, it belongs to super sound characteristics, and it can reflects the personality characteristics of the speaker from different degree; this research is one of important research in speech signal processing, now the research has broad application prospects and important theoretical significance. Prosodic features in this paper(acoustic features of time-varying evolution) mainly includes formant features, phoneme duration characteristics, pitch frequency trajectory characteristics and energy contour features, which can reflect the characteristics of the speaker’s personality.For segmental characteristics of formant parameters, we extract formant features by using the Fourier change method and Hilbert Huang transform method, and after comparing the effectiveness of the characteristic parameters, the results show that the Hilbert Huang transform method has good performance. For fundamental frequency track characteristics, analysis of variance method and circulating separation method were used to extract feature for comparison, it shows that the variance analysis method don’t need multiple threshold settings from the implementation complexity, it is not easy to appear unvoiced and voiced misjudgment phenomenon. For energy contour feature, we extract characteristic parameters of energy contour using the Hilbert Huang transform method and Fourier transform method, through the comparison of the clustering performance parameters, the results show that the performance of Hilbert-Huang transform method is better. For phoneme duration characteristics, we sentence the unvoiced voiced and silent of the speech signal by using a method based on three parameter combination,and then modeling phoneme duration parameters.Compressing the four prosodic features data, and analyzing clustering performance, then respectively calculating their clustering performance parameters. This paper establishes super sausage model based on bionic pattern recognition theory for the recognition system, according to the size of the clustering performance parameters of four parameters, we puts the four characteristics into the recognition system one by one, and then we combine them gradually into recognition system. During the individual recognition, the results show that the better the clustering effect characteristics is the higher the recognition rate is, so the choice of reasonable and effective features can improve the system recognition time. During the features combined recognition, the more the identification characteristics are, the clearer the coverage outline is, and then the correct recognition rate is higher, the deterrent rate will be lower.