Whisper Intelligibility Enhancement Using a Supervised Learning Approach

Whispered speech can be effectively used for quiet and private communications over mobile phones. It is also the communication means of laryngectomized patients under a regime of voice rest. However, little progress has been made on the enhancement of whispered speech because of its special acoustic characteristics. Recent studies with normal-hearing listeners have reported large gains in speech intelligibility with the binary mask approach. This method retains the time-frequency (T-F) units of the mixture signal that are stronger than the interfering noise (masker) and removes the T-F units where the interfering noise dominates. In this paper, a supervised learning method to enhance whispered speech is introduced. A binary mask estimated by a two-class SVM classifier is used to synthesize the enhanced whisper. Amplitude modulation spectrum (AMS) and frequency modulation spectrum (FMS) are extracted as input to SVM. Speech corrupted at low signal to noise (SNR) levels with different types of maskers is enhanced by this method and presented to normal-hearing listeners for word identification. Experimental evidence from the listening tests indicated substantial improvements in intelligibility over that attained by human listeners with unprocessed stimuli.