An Acoustic Modeling Method Robust against Changes of Speaking Style in Error Recovery

Synopsis The performance of automatic speech recognition system is not enough to perfectly recognize speech and users must rephrase their input utterances when recognition errors occur. However, speaking styles of error recovery utterances are significantly changed from that of normal utterances. Because of these changes, the performance of speech recognition system is degraded to the contrary. In Japanese, especially, the occurrence of syllable-stressed utterances increases in error recovery. In this paper, we propose an acoustic modeling method that is robust to syllable-stressed speech. In syllable-stressed utterances, each syllable is uttered in a manner between a continuous utterance and an isolated syllable utterance and the continuity of each syllable is changed. To handle the acoustic characteristics of syllables uttered like an isolated syllable, we use vowel triphone models taken from existing acoustic models, which are succeeded by silence. To handle the change of continuity of each syllable, we use left-context dependent vowel biphone models, which are generated by using the same training data as that used for creating the existing acoustic models. These models and conventional triphone acoustic models compose a single acoustic model by taking a multi-path approach. Our method improves the performance of speech recognition systems for syllable-stressed speech without any need to collect additional training data. In addition, it provides a slightly better performance for normal utterances.