Is it possible to add another language?
As I see from description, it supports only English right now. Can it be fine tuned for another language? If yes, then what is the recipe?
@zaskara I couldn't find a published recipe for fine-tuning Phonon-2 itself. Two routes, depending on your language:
One of Parakeet v3's 25 languages: you may not need a fine-tune. Phonon-2 is derived from nvidia/parakeet-tdt-0.6b-v3, which transcribes bg, hr, cs, da, nl, en, et, fi, fr, de, el, hu, it, lv, lt, mt, pl, pt, ro, sk, sl, es, sv, ru and uk, and detects the language itself. The full NeMo checkpoint download is about 2.5 GB; Phonon-2's packed download is about 164 MB.
Any other language: that card says the base model "can be used as a pre-trained checkpoint ... for fine-tuning on another dataset" in NeMo (pip install -U nemo_toolkit['asr']).
Phonon-2 itself looks harder. Its NOTICE says the encoder's linear weights were re-trained and quantised to a five-value alphabet, and lists CHiME-6 and SPGISpeech as training audio. The repo ships a packed inference container. I found no training code in it or in the phonon docs. I can't confirm a supported way to fine-tune Phonon-2 itself from those sources.
I haven't run either model on non-English audio. This is from the model cards and NOTICE only.