Lexicon Optimization for WFST-Based Speech Recognition Using Acoustic Distance Based Confusability Measure and G2P Conversion

In this paper, we propose a lexicon optimization method based on a confusability measure (CM) to develop a large vocabulary continuous speech recognition (LVCSR) system with unseen words. When a lexicon is built or expanded for unseen words by using grapheme-to-phoneme (G2P) conversion, the lexicon size increases since G2P is generally realized by 1-to-N-best mapping. Thus, the proposed method attempts to prune the confusable words in the lexicon by a CM defined as an acoustic model distance between two phonemic sequences. It is demonstrated by LVCSR experiments that the proposed lexicon optimization method achieves a relative word error rate (WER) reduction of 14.72% in a Wall Street Journal task compared to the 1-to-4-best G2P converted lexicon approach.