Mixing language

[TaciteOFF]

Hello,

I used it with parakeet-tdt-0.6b-v3 to transcribe a radio show but as soon as it detects one or two words in english, the whole sentence becomes an english sentence.
Is there a way to force a language ?

Thank you.

[senstella]

Unfortunately, not that I'm aware of! The model inherently doesn't seems to support code switching. Simply sampling from the language you want is likely to degrade the performance and would not output the English tokens at all, which can be problematic for some languages. However, Canary model is likely to support initial language settings as it has Transformer decoder, and language token in the prompt (like Whisper). Though I'm still working on it at #30!

It might be worth asking about this in the model discussions here:
https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3/discussions/

[TaciteOFF]

Thanks :)

[matijagrcic]

For more context:

https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3/discussions/8

output = asr_ast_model.transcribe(['2086-149220-0033.wav'], source_lang='en', target_lang='fr')
print(output[0].text)