Noise spectrum estimation with improved minimum controlled recursive averaging based on speech enhancement residue

摘要

The conventional soft-decision based noise estimation algorithms normally assume that noise exists, only when speech is absent. Consequently, the estimated noise spectra are not updated in the segments of speech presence, but only in those of speech absence. This assumption often results in several problems such as delay and bias of noise spectrum estimates. In this paper, we propose a solution by using speech enhancement residue (SER) to compensate the estimation bias in the presence of speech. The proposed method can be naturally combined with the improved minimum controlled averaging (IMCRA) method to consistently update noise spectra. The experimental results show that the SER-based IMCRA can reduce the relative segmental estimation errors for various types of noise at different SNR levels, especially for car internal noise.