A speech enhancement algorithm based on computational auditory scene analysis

Research on the ETSI speech enhancement system was conducted using traditional Wiener filter for noise reduction, which performed well when signal-noise ratio was high enough. However, when SNR decreased to a certain extent, it failed to suppress pulse noise effectively. Computational auditory scene analysis (CASA) simulating human auditory characteristics could make up for this weakness. Therefore, based on ETSI combined with CASA, a new speech enhancement algorithm was proposed, which performed feature extraction and spectrum estimation in the Gammatone domain rather than the original Mel domain as well as filtered out noise by an ideal ratio mask (IRM). On the noisy subset of the TIMIT corpus, the proposed enhancement achieves higher objective acoustic quality and proven ability to inhibit pulse noise under low SNR conditions compared to the original system. It also obtains an improvement in terms of the reduction of word error rates under low SNR conditions in the back-end speech recognition system.