Enhancing speech intelligibility in reverberant spaces by a speech features distributions dependent pre-processing

In this paper, we deal with a pre-processing based on speech envelope modulation for intelligibility enhancement in reverberant large dimension public enclosed spaces. In fact, the blurring effect due to reverberation alters the speech perception in such conditions. This phenomenon results from the masking of consonants by the reverberated tails of the previous vowels. This is particularly accentuated for elderly persons suffering from presbycusis. The proposed pre-processing is inspired from the steady-state suppression technique which consists in the detection of the steady-state portions of speech and the multiplication of their waveforms with an attenuation coefficient in order to decrease their masking effect. While the steady-state suppression technique is performed in the frequency domain, the pre-processing described in this paper is rather performed in the temporal domain. Its key novelty consists in the detection of the speech voiced segments using a priori knowledge about the distributions of the powers and the durations of voiced and unvoiced phonemes. The performances of this pre-processing are evaluated with an objective criterion and with subjective listening tests involving normal hearing persons and using a set of nonsense Vowel–Consonant–Vowel syllables and railway station vocal announcements.