A phase-aware single channel speech enhancement technique using separate bayesian estimators for voiced and unvoiced regions with digital hearing aid application

The modern digital hearing aids suffer from the problem of degradation of perceived signal quality and intelligibility of audio signal when signal to noise ratio (SNR) of the input signal becomes very low. In this work, a speech enhancement algorithm is proposed where the magnitude spectrum of clean speech is estimated by using two separate Bayesian estimators derived from two different cost functions. For the voiced regions, the perceptually motivated adaptive β-order weighted minimum mean square error (MMSE) estimator is used and the phase structure of the voiced region is also reconstructed using a harmonic model of speech. On the other hand, for the unvoiced segments, Bayesian estimator with modified Itakura-Saito (MIS) cost function is utilized for estimating the magnitude spectrum of clean unvoiced signal. The proposed algorithm is simulated under different non-stationary noisy environments at various signal to noise ratio values. The experimental results show that the proposed speech enhancement framework performs better than other standard benchmark methods in terms of several quality and intelligibility assessment metrics of perceived audio signal.