Studies On Speech Feature Parameters For Speaker Based On Wavelet Transform

Speaker recognition (SR) aims to identify or verify a person's identity by extracting the speaker individuality from a present sample utterance using signal measurement techniques in a range of registered persons. According to the difference of recognition mode, it can be divided speaker identification (SI) and speaker verification (SV). One of the most important problems in a speaker recognition system is that how to extract the appropriate and stable characteristic features of speech that can represent a speaker, which directly reflect the system's ability.STFT (short-time Fourier transform) is a traditional method in signal analysis and process, and most of common speech features are extracted by it. Wavelet transform is a wonderful method, which have adjustable resolution in time and frequency fields leading to more subtly analysis for a signal segment. It works in a good pattern according with the rule of human ears distinguishing frequencies from voice. For STFT having inevitable disadvantages in analysis of unstable signal such as speech signal, as a result of study on wavelet theory and speaker recognition techniques, two feature parameters, IWPTC (Incomplete Wavelet Packet Transform Coefficients) and WPTC (Wavelet Packet Transform Coefficients), are got based on wavelet transform. They derive from these theories including wavelet analysis, wavelet multi-resolution analysis and wavelet packet analysis, also a conventional feature parameter, MFCC (Mel frequency cepstral coefficients), which based on human auditory mechanism, for reference. A speaker identification system, which is built by Matlab tools and can identify a speaker whether in the mode of text-dependent or in text-independent, shows that the efficiency using these two new features is higher than that using MFCC feature parameters. Theory and experiment all testify that the recognition performances of these new feature parameters in SR system extracted by wavelet analysis method are better than that of features extracted by short-time analysis method.