Abstract: Aiming to the problem of speech enhancement from colored noise, the Chinese restaurant process mixture model (referred to as CRPMM) was introduced to modeling the distribution of noise. Markov sampling chains can be conveniently extended, where the latent variables are distributed according to a Chinese restaurant process. The updating form of joint probability of parametric variables and the latent variables were built related to block sampling. This method can approximate the posterior distribution of noise on space exactly, incorporating Kalman filtering technique. The simulation results and the experimental result of speech enhancement show that the infinite block sampling technique data-driving can adapt new modes, and a higher efficiency can be achieved for speech enhancement.  

Key words: speech enhancement, colored noise, Bayesian inference, Chinese restaurant process mixture model, Markov chain Monte Carlo (MCMC) sampling