Pronunciation Scoring for the Hearing-Impaired

Automatic assessment of articulation and prosody is an important aid for speech theraphy and language education. In this paper, we focus on speech theraphy for hearing-impairment and propose methods for automatic articulation and prosody scoring. The pronunciation problems of the hearing-impaired are briefly discussed. Three methods are developed for automatic pronunciation scoring for the hearingimpaired. In the first method, Hidden Markov Model (HMM) based phoneme recognition and forcedalignment is employed for detecting articulation problems. The second method performs prosody scoring by operating at the syllable level and comparing the pitch and energy contours of a given utterance with a reference utterance. Finally, speech-to-text alignment information is used for duration scoring of a test utterance with a reference utterance. The proposed methods are evaluated on a database collected from native Turkish speakers. The articulation scoring method improves the correct recognition rate of confusable pairs of words by 5.6% yielding a correct recognition rate of 84.8%. The results also show that the prosody and duration scoring algorithms provide useful information to assess the match between the prosodic characteristics of two speakers.