Cross language transcriptions

[luigisaetta]

In these days I have tested Whisper extensively and, since I'm Italian, I have also tested the capability to do cross language transcriptions (example: I submit an audio speech in Italian and I get the English transcription).

I'm really impressed by the quality of results. With my voice, it simply works out of the box!

Obviously, the contrary (speech in English, transcription in Italian) doesn't work really well, but it is something described in the model card (cross language trained extensively only for translation in English).

And I'm doing something more: I'm doing fine-tuning on a "special" english domain language. Using the medium model.
And the capability to translate general speech in Italian is not lost even if I do fine tuning, improving greatly (from 0.7 to 0.3) the WER on that special domain language.

@jongwook well, it seems to me that OpenAI guys have opened a new era. Thanks.