On automatic recognition of spectrally reduced speech synthesized from amplitude and frequency modulations

This report investigates the behavior of automatic speech recognition (ASR) system with spectrally reduced speech (SRS) synthesized from subband amplitude modulations (AMs) and frequency modulations (FMs). Acoustic analysis shows that the resynthesis of SRS from only AM components helps alleviate certain non-linguistic iabilities in the original speech signal. When the SRS spectral resolution is suciently good, this alleviation not only has no consequence but also yields comparable or even better ASR word accuracy compared to that attained with original clean speech signal. In contrast, FM components support human speech recognition but yield no significant improvement in terms of ASR word accuracy when the SRS spectral resolution is sufficiently good