whisper-ja-22M

Whisper finetune for Japanese focused on general/anime domains. For usage instructions follow openai/whisper-large-v3-turbo. Due to vocab changes ctranslate2>=4.7.1 required for faster-whisper. For inference engines with hardcoded vocab, the token embedding can be padded.

Finetuned from tiny with pruned vocab and encoder conv adaption, indices can be found in mapping.txt. Trained decoder only for 2^19 steps, batch size 64. Using a 45000 hour corpus (largest source is 17000 of filtered reazonspeech-all) with custom mixing ratio and augmentation to maintain long form performance and timestamps.

Benchmarks. CER roughly between OpenAI whisper-base/small, not great but also the smallest model I've tested. If you can run something bigger I would suggest moonshine-tiny-ja (27M), whisper-ja-51M, or a larger NAR model. Depending on inference engine the speed difference can be minimal because of overhead. Also trained for lyrics but untested.

Acknowledgements

• Train sets: OOPPEENN, Reazon, 小虫哥_, Common Voice 20, deepghs
• Test sets: KitsuneX07, TEDxJP, kotoba-tech, Saruwatari-lab, grider-withourai