A study of robust speech recognition using modified running-spectrum-filtering

Abstract

This paper proposes Low-Pass running spectrum filtering (Low-Pass RSF) with low filter order. Using the proposed filtering, additive noises in high frequency band of the modulation spectrum domain are reduced and high speed calculation mechanism are realized. The cepstrum mean subtraction method (CMS) and dynamic range adjustment (DRA) are also employed with Low-Pass RSF. Using the proposed method, the similar recognition performance as conventional noise robust speech recognition can be obtained with quite low calculation cost.