Direct control on modulation spectrum for noise-robust speech recognition and spectral subtraction

On the modulation spectral domain, speech components and noise components can be separated by their modulation spectral bands. Although conventional methods extract speech components using band-pass filtering on running spectrum, those extractions are affected by limitation of filtering such as stability, filter steepness, etc. Instead of it, this paper proposes modulation spectral control which realizes the feature extraction of modulation spectral bands with weighting factors on modulation spectral domain. It is free from those limitations and enables flexible extraction of specific modulation spectral bands. The combination of it and spectral subtraction shows outstanding performance and flexibility for the variety of noise in speech recognition experiments