Research Of Speech Enhancement On Phase Space Reconstruction

Speech will be polluted by noise when being processed or transmitted. People have more and more requests on speech quality today. Speech enhancement is a subject about how to reduce noise component in the polluted speech, avoid distortion and make the signal more pellucid and comfortable as far as we can.Chaos is a subject concerning evolution and its intrinsic orders in a system. It reveals the orders behind apparently chaotic movement and provides us a new way to understand complicated phenomenon, so it is widely used in many fields. The foundation of chaotic time series analyse is phase space reconstruction, which changes single variable series to multi-dimension space, so we can analyze time series in chaotic theories, and reveal the orders in it.In this thesis, the chaotic characteristics in Chinese speech are analyzed based on phase space reconstruction theory. Parameters such as Largest Lyapunov Exponent and Correlation Dimension are calculated by programming. We can see the ample chaotic information in Chinese speech, and Chinese speech is a non-linear system between deterministic periodic signal and indeterministic stochastic signal. Correlation Dimension is an important parameter to describe the construction of attractors in chaotic signal. Correlation Dimension in speech is greatly different from that in noise, which is the theoretic premise for distinguishing useful signal from noise in speech enhancement.In this thesis, we will apply phase space denoising to speech enhancement. Speech signal concentrates in a subspace of finite dimensions, but noise spreads to all directions. After reconstruction, we can reduce noise by finding signal subspace and removing noise subspace. In this thesis, the method of looking for neighbor points in phase space was improved to simplify calculation. In the experiments, Chinese phonemes were enhanced by this method, and were evaluated in an objective way. Continuous speech was evaluated in both subjective and objective ways. From the experiment results, we can see that phase space speech enhancement has considerably satisfying effect for higher SNR and less distortion.