Enhancement techniques to improve the intelligibility of consonants in noise : speaker and listener effects

The aim of our work is to increase the intelligibility of speech in noise by modifying regions of the signal that contain acoustic cues to consonant identity in order to make it more resistant to subsequent degradation. Two instances of each of 36 vowelconsonant-vowel (VCV) stimuli comprising the consonants D F I R V M H X U \ O P in the context of the vowels # K W were recorded by two male and two female speakers without any phonetic training. These tokens were manually annotated; the vowel onset/offset and consonantal constriction/occlusion regions were then selectively amplified, combined with speech-shaped noise at 0 dB SNR and presented to a group of 14 native-English listeners. Significant increases in intelligibility between the natural and enhanced conditions were obtained for all speakers but the extent of the improvement was greater for the initially least intelligible speakers. In a second experiment, speech material for two of the four speakers was presented to three new groups of native English, native-Japanese and native-Spanish L2-learners of English. For all groups, consonant intelligibility was significantly higher in the enhanced condition. The extent and patterns of errors were related to the ‘distance’ between the phonological systems of the listeners’ L1 and L2 for the set of consonants under investigation. Results of these two experiments demonstrate the robustness of our enhancement techniques across speaker and listener types.