Speech Enhacement Based On Kalman Filtering

In the noise environment,the performance of most speech processing systems deteriorates sharply. As a solution to noise pollution, speech enhancement is an effective technology, and has been being the research focus of the speech signal processing all the time. The purpose of speech enhancement is to extract clean speech signal from the noisy speech as far as possible, to improve signal to noise ratio (SNR) and speech quality.Kalman filter is an optimal linear estimator in the minimum mean square error criterion, with non-stationary signal processing capacity. Speech enhancement based on kalman filtering, integrating with speech generation model, can be applied in non-stationary noise environment. In this thesis, speech enhancement based on Kalman filtering is in-depth studied and the following work has been done:1. Predictor, filter and smoother in kalamn filtering theory and their application in speech enhancement technology is studied. A speech enhancement system based on kalman filtering is realised.2. There is much much residual noise in the enhanced speech based on kalman filtering, the thesis introduces a method to reshape noise power spectrum through speech spectral likelihood ratio. Minimum statistical tracking is a common noise power spectrum estimation method, but the estimated value is often low. So the thesis introduces a method to reshape noise power spectrum through speech spectral likelihood ratio. It increases the estimation of noise power spectrum in the frequency components where speech is weak. Combined with this method, the speech enhancement based on Kalman filtering can reduce residual noise significantly, and make enhanced speech clearer and more natural.3. By using the subband decomposition techniques, we propose a subband speech enhancement method based on Kalman filter. Experiments results show that this method improves the quality of the enhanced speech and also largely reduces the computation complexity due to the low orders of models in subbands, and thus it can be easily realized real time.