Noise model adaptation in model based speech enhancement

This paper presents a noise model adaptation algorithm for model based speech enhancement (MBSE). Noise model adaptation is essential for proper operation of MBSE in non-stationary noise environments. The proposed algorithm updates the model to reflect changes in the amplitude, spectral shape and sources of the noise. Noise model codewords are selected for retraining based on a distance measure from noise-only input vectors, and a new centroid is formed based on a moving-average window length which determines the adaptation characteristics. An efficient algorithm for updating the probability structure of the noise model is presented. The adaptation algorithm is evaluated and shown to improve the performance of minimum mean-square error MBSE against actual non-stationary noise environments