Add support for new models

[istupakov]

Parent issue for requests and ideas for adding new models.

[jackalxyz]

@istupakov any chance to add support for https://huggingface.co/nvidia/canary-1b-flash ? It's performance is way better than canary-1b-v2 for the four common languages https://huggingface.co/spaces/hf-audio/open_asr_leaderboard

[jackalxyz]

by the way, I've tried https://github.com/istupakov/onnx-asr?tab=readme-ov-file#convert-model-to-onnx but the exported model does not work unfortunately.

[jackalxyz]

The reasoning to have canary-1b-flash because this model allows explicitly setting source and target language which yield a lot lower WER. This model's ranking (nvidia/parakeet-tdt-0.6b-v3) at ASR leadboard is high but it only autodetects input language which can get the wrong language at times

[istupakov]

by the way, I've tried https://github.com/istupakov/onnx-asr?tab=readme-ov-file#convert-model-to-onnx but the exported model does not work unfortunately.

Yes, this script is only for Conformer/Parakeet models. Converting Canary models is more tricky.