Multiband Excitation for Speech Enhancement

The speech enhancement algorithm proposed aims to improve the quality of denoised speech by introducing voicing information to a Wiener-type spectral amplitude gain function. A constrained multiband excitation (MBE) model is used to emphasize harmonic components of the glottal input; and a low variance and bias autoregressive multitaper (ARMT) estimate models the vocal tract. Objective and subjective evaluations show improvement over unconstrained models and those using high variance spectrum estimators.