Speech enhancement employing adaptive boundary detection and morphological based spectral constraints

A speech enhancement algorithm which employs noise adaptive boundary detection and morphological based spectral constraints is presented. The technique is formulated in the frequency domain, and uses a speech-specific weighted subtraction factor and power exponent, followed by the application of morphological based smoothing constraints. Detected speech boundary information allows the enhancement procedure to adapt and thereby track changing speech characteristics. Morphological operators based on Minkowski (1903) set operations of addition and subtraction allow for smooth spectral transitions during phonation. Results from speech degraded by additive Gaussian noise show that the algorithm: provides superior speech quality over all speech sound classes; time-versus-LPC spectra plots show improved vocal tract characterization; and the enhancement algorithm produces higher quality versus SNR when compared to more traditional methods such as standard spectral subtraction and short-time Wiener filtering