Acoustic Phonetic Decoding for assessment of mispronunciations in speakers with cognitive disorders

This paper1 introduces the results in
acoustic-phonetic decoding over a corpus of isolated
words uttered by young speakers with speech
impairments due to cognitive disorders. The low
performance in Phone Error Rate (PER) in this task
does not mean that it is an imprecise system, because
speakers are producing more than 17% of mistakes
and deletions in their pronunciation. Hence, it is
shown how the outcome of the acoustic-phonetic
decoder is matched to the actual pronunciation of
the speakers, resulting in a detection performance of
15% in False Acceptance Rate (FAR) with a 20% of
False Rejection Rate (FRR) when using speaker
dependent acoustic models. The comparison
between a rule-based grammar obtained from the
rules of syllable construction in Spanish and a data-
driven grammar learned from a large Spanish text
corpus show no difference in the acoustic-phonetic
decoding task and in the mispronunciation
assessment.