Coarticulation across communicative contexts: An acoustic analysis of the LUCID corpus using spectral and temporal measures

Abstract

Coarticulation reflects a balance between talker efficiency and speech comprehension. This study examines coarticulation across different communicative contexts in the LUCID corpus. The talkers completed an interactive spot-the-difference task with a listener when there were no communication barriers (NB), when their speech was vocoded (VOC), when it was masked by multi-talker babble (BABBLE), and when the listener was an English language learner (L2). They also read sentences without an interlocutor conversationally (READ-CO) and clearly as if talking to someone with hearing loss (READ-CL). The spectral distance and the relative duration of coarticulatory transition between the adjacent segments were computed. Both measures showed that speech produced in the presence of a communication barrier (BABBLE, VOC, L2, and READ-CL), be it real or imagined, is more resistant to coarticulation than speech produced in the barrier-free conditions (NB and READ-CO). Spectral distance was more sensitive to differences among all conditions involving a real listener (BABBLE, VOC, L2, and NB) while relative transition duration measure better captured differences between read and spontaneous speech for the barrier-present conditions (BABBLE, VOC, L2, and READ-CL). The results suggest that talkers reduce coarticulation in response to the particular communication challenges which may aid speech perception.