Improvement of CRF-Based Accent Sandhi Prediction Using The Features Derived from Accent Rules

When developing Japanese text-to-speech (TTS) systems, algorithms to accurately predict accent types of each constituent phrase is essential for better output speech quality. In our previous studies on the accent type estimation, a CRF-based method was realized. Although this method outperformed the conventional rule-based method, the estimation accuracy of particular phrases such as those including numerals or loanwords was still not sufficient. In this paper, we newly added the features used in the rule-based estimation of these phrases as CRF features. The experimental result for JNAS corpus showed improvements in accent type estimation. As an example of possible applications of the developed method other than speech synthesis, we constructed an accent type prediction module for CALL systems. This module can automatically generate accent dictionaries of conjugation words for any Japanese texts.