Use just released Whisper 3 Turbo for 5 times faster inference

[menelic]

Open AI just released Whisper 3 turbo: https://huggingface.co/deepdml/whisper-large-v3-turbo

According to several evaliations, it offers up to 5x faster inference speed with only a minimally higher error rate in some cases: https://huggingface.co/openai/whisper-large-v3/discussions/160

Please integrate it in neScribe, a higher speed of transcription would really expand and deepen use cases for this great software.

[kaixxx]

Nice catch, I wasn't aware of the new model. Looks very promising. It may introduce a little more hallucinations on noisy audio, a problem that is already known from the normal large-v3 (for this reason, noScribe still uses large-v2). But the speed improvements might still justify a switch. I will test it.
I have to update some libraries and test everything, so please be patient. But I will try it out for sure. I'll update this issue if I have any results.
Thanks!

[wwwebweber]

For Swiss German, large-v2 remains the best. v3 and v3-turbo have forgotten Swiss German.

[kaixxx]

For Swiss German, large-v2 remains the best. v3 and v3-turbo have forgotten Swiss German.

Ah, interesting observation. Generally, there have been a lot of controversy if v3 is really an improvement or made things worse. Benchmarks are not everything.
I think we should implement a way for users to download and install custom models into noScribe. This would also allow using fine-tuned models for certain languages etc.
For instance, there is a model that transcribes Swiss Dialect in verbatim form, not 'translated' into German: https://huggingface.co/ss0ffii/whisper-small-german-swiss Unfortunately, it's based on the small whisper model which gives only mediocre quality. But it's interesting nonetheless.