Automatic rule generation for linguistic features analysis using inductive learning technique: linguistic features analysis in TOS drive TTS system

The linguistic features analysis for input text plays an important role in achieving natural prosodic control in text-to-speech (TTS) systems. In a conventional scheme, experts refine suspicious if-then rules and change the tree structure manually to obtain correct analysis results when input texts that have been analyzed incorrectly. However, altering the tree structure drastically is difficult since attention is often paid only to the suspicious if-then rules. If earlier rule-tree structure is inappropriate, any attempt to improve the performance may be limited by the stiffness of the structure. To cope with these problems, the new development scheme generates analysis rules by using C4.5 [1], where an if-then rule-tree structure is generated by off-line training. The scheme has the advantage that since the generated rule-tree structure is simple, the rules are easier to maintain. The scheme is applied to generating four types of analysis rule-trees: rules for forming accent phrases, rules for determining accent position, rules for analyzing syntactic structure, and rules for pause insertion. An experimental evaluation was performed on these four rules. The accuracy was 96.5 percent for the accent phrase formation, 95.5 percent for the accent positioning, 87.0 percent for the pause insertion, and 88.3 percent for the syntactic analysis despite using small training data. These results indicate the validity of the scheme. The new scheme is used for developing linguistic features analysis rules in a Japanese TTS system, TOS Drive TTS [3].