Computationally Efficient Speech Enhancement By Spectral Minima Tracking In Subbands

We present an efficient algorithm for the enhancement of speech signals which are heavily corrupted by short-time stationary, acoustically or electrically added disturbances. The algorithm is based on spectral amplitude estimation using an overlapadd FFT filter bank system. Compared to other systems, the improved performance of our speech enhancement system is achieved by the combination of the best known spectral amplitude estimators of the noisy speech signal and a new efficient and reliable noise spectrum tracker. As a result, our speech enhancement system requires no speech pause detection for noise estimation and needs only 14% -- 23% of the resources of a commercially available digital signal processor. I. Introduction T HE enhancement of noisy speech is a challenging research field with applications including suppression of environmental noise in machinery halls, mobile radio communications systems, noise suppressors for automatic speech recognition systems, and hearing aids...