The Research Of Time-frequency Modeling For Speech Signal

Speech signal is a non-stationary time-varying signal, carrying all kinds of information. Modeling accurately is important to subsequent processing such as speech coding, speech compression, speech recognition.Time-frequency analysis is an effective tool for non- stationary signal analysis,one as a new research area of signal processing in recent years,can analyze signal in both time domain and frequency domain and offer a new theory and method for speech modeling ,so it has theoretical significance and application value to discuss speech modeling.Time-frequency modeling for speech signal is studied in this paper, and the research work is as follows:Firstly, a sinusoidal harmonic spectrum model based on wavelet multi-resolution analysis is proposed.Sinusoidal harmonic spectrum model has the problem of fixed frame length, which can not get the best resolution of each harmonic, and the resolution determines the effect of speech modeling. The combination of wavelet transform and sinusoidal harmonic spectral model has multi-resolution characteristics, which solves the problem of fixed frame length,decomposing an input speech signal into multi-resolution sub-band signals using the wavelet transform, and sinusoidal harmonic spectrum speech model is applied to each sub-band signal respectively,finally,each sub-band signal modeled is synthesized,which called time-frequency modeling.The simulation experiments show that the signal reconstruction error of the proposed model is reduced by about two orders of magnitude,and MOS's grades have increased about 0.3 through PESQ testing.Secondly, time-frequency analysis of speech signal based on improved S transform is discussed. S transform is a combination of short-time fourier transform and wavelet transform, which has the advantages of the both and has broad prospects. Select the appropriate mother wavelet according to speech signal characteristics to improve the S transform and improved S transform can used to analyze speech signal and extract the parameters of the speech signal, Simulation experiments show the time-frequency structure of the improved S transform is finer compared with STFT and wavelet transform.Finally, the transient time-frequency modeling of speech signal by matching pursuit with a wavelet-based dictionary is proposed. The core idea of matching pursuit is searching for the best matching atoms, and after building wavelet-based dictionary with daubechies wavelets, matching pursuit is used to find the best wavelet atom, and then speech signal is reconstructed by a series of the best wavelet atoms. Simulation experiments show that the proposed model is more effective and superior than matching pursuit with sinusoidal dictionary.