Optimal noise estimation under speech presence uncertainty using minimum Statistics Method

Frequency domain speech enhancement algorithm used in the single-microphone mobile phone usually needs a noise estimator. Minimum Statistics Method can track noise robustly. However, many algorithms based on minimum statistics method do not optimize their tracking factor. In this paper, we derive a new global optimal noise tracking factor controlled by another two new parameters. By combining with speech presence uncertainty and minimizing the conditional mean square error, we obtain a global adaptive factor that does not require error monitor, and then we integrate it with minimum statistics method to estimate noise. Finally, the proposed algorithm is evaluated by using it in Log-Spectral Noise Suppressor and results show a better performance.