Global soft decision based speech enhancement using voiced-unvoiced uncertainty and harmonic phase decomposition technique

This paper introduces a single-channel speech enhancement framework based on global soft-decision, where the magnitude spectrum of clean speech is estimated by combining two separate Bayesian estimators based on voiced-unvoiced uncertainty. For the voiced regions, the perceptually motivated adaptive β-order weighted minimum mean square error (MMSE) estimator is employed. On the other hand, for the unvoiced segments, Bayesian estimator derived from modified Itakura-Saito (MIS) cost function is utilized. In addition to this spectral magnitude enhancement, the clean speech is reconstructed in time domain using the estimated clean phase based on a technique involving harmonic phase decomposition and spectro-temporal smoothing filters. The proposed algorithm is simulated under different non-stationary noisy environments at various signal to noise ratio values. The experimental results show that in contrast to other benchmark methods, the proposed speech enhancement method produces enhanced speech signal with negligible musical noise and a joint improvement in both perceived quality and speech intelligibility.