Reverberant speech enhancement by spectral processing with reward-punishment weights

This paper presents an approach for the suppression of late reverberation and additive noise in single-channel speech recordings by spectral processing with reward-punishment weights. The spectral variance of the late reverberant signal can be estimated directly from the received reverberant signal using a statistical reverberation model and a limited amount of a priori knowledge about the acoustic channel between the source and the microphone. However, the suppression of late reverberation by spectral subtraction tends to degrade disproportionally low-level signal regions and signal transients. In this work, several reward-punishment criteria such as correlation parameter, Spectral Flatness Measure (SFM), Peak-to-Sidelobe Energy Ratio (PSLER), are taken into account to avoid degrading low-level signal regions and signal transients by identifying and enhancing the high signal-to-reverberation ratio (SRR) regions in a signal-dependent fashion. The performance of our method is demonstrated by experiments using synthesized room impulse responses. The experimental results indicated that this method provides superior speech quality to state-of-the-art late reverberation suppression algorithms.