Research On Time-Frequency Modeling And Enhancement For Speech Signal Based On Matching Pursuit

In modern signal processing, Speech Signal is a typical example of non-stationary signal, characterized by a limited duration, and time varied. Joint time-frequency analysis (time-frequency analysis) is a sophisticated method to analyze its characteristics. Time-frequency analysis focused on the time-frequency characteristics of the composition of real signal. A one-dimensional signal is represented to two-dimensional time - frequency density function. In the cases of low signal to noise ratio, time-frequency analysis can also be used to achieve good results. As a new signal processing method, its application areas are more and more wide.From the view of time-frequency analysis point, in time domain and frequency domain, the speech signals and the noise have a clear difference. Inspired by this, this paper presents the time-frequency modeling method to speech enhancement. This paper introduced a local cosine-basis based dictionary, it segments the time frame and adapts speech signal's changes of the time-frequency structure. We also get a higher frequency resolution. To break the constraint of uncertainly principle in traditional Spectrographic Analysis method, we propose using the combination of local cosine-basis atoms'WVD to analyze the speech's Time-Frequency structure. This method has high Time-Frequency resolution as WVD but the cross-term interference in WVD. Through the analysis between time and Time-Frequency plan, we proved that it is reasonable to use local cosine-basis dictionary to model speech signal.In all kinds of speech enhancement arithmetic, how to get time-frequency model parameter extraction is a difficult problem. We use matching pursuit decomposition method to get time-frequency model parameter extraction. Choosing a suitable dictionary, to decompose noised signal, the noise in the time-frequency plane is diluted, and speech component is comparatively gathered in the time-frequency plane in a given region. In each step, we will extract the parameters of atoms which are the most relevant with speech signal. The simulation results show that, matching pursuit method is auto-adapted. The time-frequency characteristics of the speech signal can be good gathered in the atoms in the dictionary.Matching pursuit decomposition can control the iterative number to achieve the purpose of filtering noise. But when the signal to noise ratio is too low, the energy of noise atoms may be higher than the speech atoms', this will break down the noise atoms as speech atoms. To this end, this paper presents matching pursuit decomposition combines with sub-space method. The vector space of noisy speech is considered as a noisy speech space added a pure nose space. We can use of sub-space decomposition technology to remove pure noise sub-space. Simulation results show that the match pursuit with the sub-space, applies to the various SNR speech enhancement.