From stochastic speech recognition to understanding: an HMM-based approach

This paper presents results achieved with an HMM-based stochastic speech understanding approach. The applied embedded training is directly adapted from continuous speech recognition and utilises transcribed text corpus without explicit time alignments. The proposed method is tested on two databases, one in English, the other one in Finnish, from two different demonstration applications (currency inquiry and city bus timetable inquiry systems). The results indicate the applicability of the proposed method and show that semantically relevant parts of input queries can be identified with a 5-8% error rate on the semantic unit and 13-20% error rate on the sentence level. The segmentation capability of the approach indicates that the system is capable of exploring the meaningful parts of the queries in an unsupervised fashion