support transcription from microphone and other live sources + translation

[KonradHoeffner]

noScribe works extremely well if we have prerecorded audio but sometimes we have live presentations that we want to transcribe and translate on the fly, is there a chance that this will be supported in the future or is that totally out of scope?

While there are other tools like https://github.com/chidiwilliams/buzz for that and usually tools are best that have a small and focused goal, I want to argue that this might be a different case:

1. Behind both noScribe and buzz there is the same underlying Whisper model which offers this functionality.
2. Both tools have large download and install sizes, which means if one needs one and only has the other it means both a long download time and also a lot of space on the storage medium.
3. My colleagues and I find buzz quite difficult to install.

[kaixxx]

I'll have to think about this. The transcription itself with whisper might work. But we would need to add a lot of supporting infrastructure to noScribe: the ability to select a microphone and record audio, buffer audio, and stream it to a file (because I wouldn't be happy with not having an audio recording to compare the final transcript against). Also, speaker detection would not work with this setup.

[KonradHoeffner]

Thanks for the swift response! If it's too much work and too far from the core mission of this tool, would another option be to separate the models from the tool itself so they can be shared?
As tools from GitHub are usually are in the megabyte-range rather than several gigabytes for Buzz and noScribe, I suspect most of that storage is used by Whisper, is that correct?
If yes, would it be possible to just share the Whisper model and thus drastically cut down on the download size?
Similar to how one only has to install ffmpeg or Visual C++ redistributables or an audio or video codec or the Java Virtual Machine only once.