Nonparametric Bayesian dereverberation of power spectrograms based on infinite-order autoregressive processes

摘要: This paper describes a monaural audio dereverberation method that operates in the power spectrogram domain. The is robust to different kinds of source signals such as speech or music. Moreover, it requires little manual intervention, including complexity room acoustics. based on non-conjugate Bayesian model spectrogram. It extends idea multi-channel linear prediction domain, and formulates reverberation non-negative, infinite-order autoregressive process. To this end, interpreted histogram count data, which allows nonparametric be used prior for process, allowing effective number active components grow, without bound, with data. In order determine marginal posterior distribution, convergent algorithm, inspired by variational Bayes method, formulated. employs minorization-maximization technique arrive at an iterative, algorithm approximates distribution. Both objective subjective evaluations show advantage over other methods spectrum. We also apply music information retrieval task demonstrate its effectiveness.