Impact of frame rate on automatic speech-text alignment for corpus-based phonetic studies

Phonetic segmentation is the basis for many phonetic and linguistic studies. As manual segmentation is a lengthy and tedious task, automatic procedures have been developed over the years. They rely on acoustic Hidden Markov Models. Many studies have been conducted, and refinements developed for corpus based speech synthesis, where the technology is mainly used in a speaker-dependent context and applied on good quality speech signals. In a different research direction, automatic speech-text alignment is also used for phonetic and linguistic studies on large speech corpora. In this case, speaker independent acoustic models are mandatory, and the speech quality may not be so good. The speech models rely on 10 ms shift between acoustic frames, and their topology leads to strong minimum duration constraints. This paper focuses on the acoustic analysis frame rate, and gives a first insight on the impact of the frame rate on corpus-based phonetic studies.