AR Model-based Bayesian Speech Enhancement for Nonstationary Environment

A new technique for enhancing audio signal from a noisy nonstationary environment is presented in the paper. Autoregressive (AR) model is used to efficiently exploit the temporally correlated information of audio and noise signals during a short stationary frame. The temporal models of signals and noisy process are combined to construct a state space. The state space appropriately describes that the observed noisy signal is generated from two underlying sources which evolve with Markovian dynamics across successive step times. In the state space, the clean speech and the noise are two hidden source signals. The recovery of clean speech and the estimation of all the model parameters are carried out within the variational Bayesian framework. The original speech can be estimated as a state using a variational Kalman smoother. The experimental results show that our approach can obtain better performance in terms of signal-to-noise ratio (SNR) measure.