Integrating the cochlea's compressive nonlinearity in the Bayesian approach for speech enhancement

The human ear has a great ability to isolate speech in a noisy environment and, therefore, constitutes a great source of inspiration for speech enhancement algorithms. In this work, we propose a Bayesian estimator for speech enhancement that integrates the cochlea's compressive nonlinearity in its cost function. When compared to existing Bayesian speech enhancement estimators, the proposed estimator can achieve a better compromise between speech distortion and noise reduction by favoring less speech distortion at lower frequencies, where the main formants are located, while increasing the noise reduction at higher frequencies. The proposed estimator also yields better results both in terms of objective and subjective performance measures.