Recent Evaluations of a WFST-Based Speech Recognition Decoder

摘要

This paper describes the latest performance evaluations on the Tokyo Tech Transducer-based (T~3) speech decoder. These evaluations focus on two particular tasks which include a large-vocabulary continuous speech transcription system with a 460k vocabulary evaluated on the JNAS corpus, and a voice search system developed for an all-Japan train timetables task. This paper provides a detailed explanation of the successful steps taken to construct a large integrated network which achieves high recognition performance, based on an exhaustive comparison of different construction strategies. Furthermore, in the context of the voice search task, this paper provides a performance comparison of two widely popular acoustic model toolkits, HTK and SphinxTrain in the unified context of the T~3 decoder. In particular these results indicate that there is a significant advantage to employing the log semiring for all WFST construction operations. These results also serve to further verify the flexibility and speed of the T~3 decoder on a variety of different tasks.