Adaptive Hidden Markov Models for noise modelling

We propose a noise estimation algorithm for single channel speech enhancement in highly non-stationary noise environments. The algorithm models time-varying noise using a Hidden Markov Model and tracks changes in noise characteristics by a sequential model update procedure that incorporates a forgetting factor. In addition the algorithm will when necessary create new model states to represent novel noise spectra and will merge existing states that have similar characteristics. We demonstrate that the algorithm is able to track non-stationary noise effectively and show that, when it is incorporated into a standard speech enhancement algorithm, it results in enhanced speech with an improved PESQ score and lower residual noise.