I have been using a local LLM called “faster-whisper-xxl” in order to transcribe a big amount of interview audio to text in Finnish. Now, after manually postcorrecting and adding scientific notations to the resulting texts, I would like to translate this text mass to English from Finnish. Faster-whisper has worked like a dream (100% recommend for transcription), but while it does have a built-in translate option, it doesn’t seem to know how to translate from text to text. Which I guess means I need another model.

Now, I would like to instruct the model, for example, to not translate and leave as-is anything that is inside brackets “[]”. Other similar character-based instructions might be needed. I would also be very happy if the model could output a colored translated text, where I could easily see the probability of correctness of each word by its color, but this is not mandatory for me. Among other things, I’d like to try and instruct it to leave commas as they are, because they now represent pauses in the original speech. The model should also not try to make grammar corrections as far as that’s possible with a translation task.

One more thing is that I need to do this using a local model without internet due to data protection, so online tools are excluded.

Now, there are hundreds of models available in Huggingface, that I can download and use, which makes it hard to decide which way to go. Can anybody suggest a model that could be good for this task? The model should be less than 12GB in size because of my device’s limitations, and naturally master the Finnish-to-English translation task.

Any recommendations? 🙂
Thank you so much again already in advance!

Yours,
Kasper