efwkjn

/

cohere-asr-ja-diffusion

like 1

Great work on the models and the benchmarks. Here is an idea you could try if you want.

The new GigaAM Multilingual foundation model (with 240M and 600M variants) claims to beat Whisper Large v3 and Omnilingual 1B on target languages under a controlled CTC fine-tuning setup. The kicker is that their 240M encoder outperforms the 1B+ baselines on average WER while being substantially smaller and faster for inference. Also, Japanese was heavily included in their Global Core pre-training cluster, so the base acoustic representation is already baked in.

Another guy recently trained a new CTC head from scratch for Bashkir on a single RTX 5090. Even though Bashkir had minimal coverage in the original pre-training data (under 500 hours), his WER dropped from an unusable 103% down to 7.2% in just over a day of training. Since Japanese has much more hours in the pretrain and since you already have clean Japanese datasets and the tokenizer pipelines ready to go, it would be sick to see if this smaller architecture can outshine 1.5B Whisper

Links:
Paper: https://arxiv.org/abs/2607.10371
GitHub with training code: https://github.com/salute-developers/GigaAM
Bashkir RTX 5090 fine-tune example: https://huggingface.co/AigizK/GigaAM-Bashkir-CV25