Underdog Woof Japanese is a small Japanese text-to-speech model. Both voices are contained in a single 40.5 MB file. The author claims it runs at approximately 14.5 times real-time speed on a single thread of an Apple M5 Max, with peak memory usage of 437 MB, and requires no GPU or API key. The author tested it on 446 Japanese sentences (including 281 sentences with difficult-to-read content such as dates, prices, phone numbers, and addresses), comparing results by kana error rate. For difficult sentences, the error rate was 1.33%, which is lower than the compared Fish Audio Drama 3, Qwen3-TTS, and Kokoro.
Why it matters It can read kanji, dates, prices, phone numbers, postal codes, and addresses using everyday pronunciation, lowering the barrier to entry for creating Japanese broadcast or learning tools.
Hugging Face likes 5 · Hugging Face trending score 5 · Hugging Face downloads 0
License: Non-commercial · ComfyUI: Not available · VRAM: Not specified
Reason
- The English version of Underdog Woof was featured in the previous Hugging Face trending list; this is the newly released Japanese version
Usage
- Run
pip install numpy pyopenjtalk-plus, download the model, then executepython -m underdog\_woof\_ja "text" hello.wav. Use--voice mikasato switch voices
Risks
- Underdog has a non-commercial license and cannot be used for commercial purposes
- The evaluation was conducted by the author; see the model card for the transcription and comparison process