The Effect of Accentuation on Vowel Recognition

This study focuses on the implementation of the phonetic effects of vowel accentuation in automatic speech recognition (ASR). The durational and spectral effects of accentuation are investigated separately by manipulating the transition and observation probabilities in hidden Markov models. We also attempt to implement the undershoot hypothesis [1], which describes spectral reduction as a direct consequence of shortening. Our findings support the widespread belief that the transition probabilities, which indirectly model durational effects, are negligible, and that the distinction between accented and unaccented vowels is determined by the observation probabilities.