Research On The Optimized Projection And Reconstruction Technology Of Speech Signal Based On Sparse Representation

In recent years, compressed sensing(CS) theory is the domestic and foreign scholars'research hot spot in signal processing fields, because its"joint sampling and compression"trait makes the A/D conversed digital sequence and conversion rate far lower than the traditional Nyquist sampling theorem, which is the most attractive. This thesis applies to CS theory in the speech signal field and especially research the optimized compression projection and reconstruction performance. This thesis analyzes the characteristics of the speech signal to optimize observation matrix so as to improve the speech signals'CS and the reconstruction performance.The author proposes a compressed sensing method of speech signal based on the optimized observation algorithm. Knowing the approximate sparsity of speech signal in the DCT domain, the method first uses the optimized observation matrix algorithm to find the optimized observation matrix corresponding to the DCT matrix and then uses the acquired matrix to project the speech signals. The experimental results show that the proposed method has better quality of reconstructed speech, so the proposed method improves the speech's CS and the reconstruction performance.Another main research content of this thesis is the adaptive CS methods of speech signal which structure the observation matrix adaptively according to the characteristics of speech signal. This thesis proposes a new inter-frame adaptive CS algorithm of speech signal, on the basis of the fact that the different types of adjacent speech frame have different inter-frame variation. The experimental results show that the proposed adaptive CS algorithm has better performance than the nonadaptive CS algorithm.This thesis further studies the adaptive CS method of speech signal and proposes a new adaptive CS method based on the judgment of unvoiced speech or voiced speech. The proposed method consists of the inter-frame adaptive step and the intra-frame adaptive step. The inter-frame adaptive step first judges that the speech frame is unvoiced or voiced speech by its short-time energy and short-time cross zero ratio. Then the step adaptively allots more observation number to the voiced speech frames according to the sum of unvoiced frames and the sum of voiced frames of the whole speech signal, because the voiced speech frames carry more information than unvoiced speech frames. After the inter-frame step determines the observation number, the intra-frame adaptive step chooses the row vectors from the completely observation matrix according to the degree of the correlation between the row vector and the speech signal. The experimental results show that the proposed adaptive CS method greatly improves the quality of the reconstructed speech. Also, the inter-frame and the intra-frame adaptive step both help to improve the CS reconstruction performance of speech signal.