Research On Subspace-based Speech Enhancement

In most applications, the aim of speech enhancement is to improve the quality and intelligibility of degraded speech, and it's important to promote the performance of speech processing system in noise environments. This dissertation mainly focuses on the subspace-based speech enhancement and the noise estimation algorithms, containing the following aspects:To minimize speech distortion and overcome the poor performance of subspace algorithms in colored noise environments, an improved optimal subspace algorithm is derived. The proposed algorithm is based on the theory of spectral domain constrained estimation and the characteristic of generalized eigenvalue decomposition. Furthermore, a unified notation of the proposed algorithm is provided for both white and colored noise environments. Results, based on informal listening tests and objective measures, indicated significant improvements in speech quality with the proposed algorithm.To make the residual noise perceptually inaudible, a well known perceptual weighting technique from speech coding is used to shape the residual noise spectrum. Under the minimum perceptually weighting errors criterion, a perceptually motivated subspace algorithm to speech enhancement is proposed in the spectral domain. The results demonstrated significant improvements in speech intelligibility with the perceptually motivated subspace algorithm.As the noise spectral estimation based on the minimum statistics introduces significant tracking latency when the noise spectral rises, an improved algorithm based on the weighted minimum statistics is proposed. Analyzing the influence of the weight on the noise spectral estimation, three kinds of typical simple curves are used to compute the weight, and the experiment shows that the weight computed by the cosine curve is the best. The simulation results showed that the improved algorithm traced the change of noise spectral quickly in most cases as well as improved the accuracy of the noise spectral estimation and the quality of speech in the non-stationary noise environment. Furthermore, to improve the performance of subspace methods in the non-stationary noise environment, the improved algorithm is extended into subspace by using the relationship between signal spectrum and autocorrelation.To estimate the noise autocorrelation from the noisy speech autocorrelation directly, an unbiased noise autocorrelation estimator is proposed under the minimum mean square error criterion, which is based on the optimal first-order smoothing recursion and energy minimum algorithm. The simulation results showed that the proposed estimator outperformed the traditional estimators, especially under the non-stationary noise environments.Finally, we apply the subspace-based speech enhancement to the speaker recognizer as the pre-processor module .The results showed that the introduced module improved the precision of the feature extraction as well as the speaker identification rates.