This pull request introduces a CMUDICT-to-Katakana dictionary, translated using ChatGPT and DeepSeek. This conversion enables the Japanese G2P mode to directly pronounce most common English words in Katakana, improving the compatibility of English terms in Japanese tts applications.
Since the translation was performed by an LLM the accuracy might not be perfect and could benefit from further refinement. However, initial testing indicates that it is currently functional and applicable in most cases.
Additionally, this update restores the "UNK" token's phoneme. During training, this token was learned as a pause, which has shown to be helpful in the Japanese G2P process.
This pull request introduces a CMUDICT-to-Katakana dictionary, translated using ChatGPT and DeepSeek. This conversion enables the Japanese G2P mode to directly pronounce most common English words in Katakana, improving the compatibility of English terms in Japanese tts applications.
Since the translation was performed by an LLM the accuracy might not be perfect and could benefit from further refinement. However, initial testing indicates that it is currently functional and applicable in most cases.
Additionally, this update restores the "UNK" token's phoneme. During training, this token was learned as a pause, which has shown to be helpful in the Japanese G2P process.