Speech enhancement based on nonparametric factor analysis

A new speech enhancement strategy is proposed by utilizing a Bayesian nonparametric method of beta process factor analysis. As a sparse representation frame work, the dictionary learning, sparse coefficients representation and noise variance estimation are integrated into a joint procedure of Bayesian posterior estimation. The beta process is adopted as a sparse prior to infer the sparsity of the signal, and the appropriate dictionary size can be inferred nonparametrically. The dictionary training process could be performed directly on the noisy speech in situ without knowing the noise variance. Experiments were executed on noisy utterances from NOIZEUS database with SNR range from 0dB to 15dB of three types of noise, like white, train and street. And the subjective and objective measures like PESQ score and the output SegSNR are implemented to evaluate the performance of the proposed method and the other state-of-the-art methods. The corresponding results show that this proposed speech enhancement method achieves better performance both in stationary and non-stationary noise conditions.