If any error occurs when switching from vosk to whisper (or vice versa) the model is set to none, but the current_backend isn't since it didn't come that far. this results in the current_backend still being vosk and the path being the same, while the model is null.

The function rec.FinalResult() that is being called in transcribe_audio.py just returns the last batch, not the whole text.
The debug logs show the whole message, but no running total is kept.

Off-topic: I have a working docker image for maubot thats compatible with whisper and vosk darmstadt/maubuntubot

I can't start the instance without vosk.
Since there is no alpine/musl version of vosk i tried to run it without but it fails when loading.

Here is the error from maubot: stacktrace.txt

And my Maubot image with the dependencies: Dockerfile.txt

Hi.

Thank you for this piece of software. It's very useful in order to transcribe into searchable text the ungodly number of Whatsapp voice messages (even work-related) one gets here in Brazil .

After living in text-based bliss for a couple of months, I noticed an issue with whispercpp. For some reason, long voice messages were not getting from ffmpeg to the model (brief messages were still getting through). The ffmpeg process would just sit there idle, presumably after completing its job, without passing the result down the pipe to whisper. After trying unsuccessfully to debug the issue for a while, I hacked the plugin to use faster-whisper instead. This seems to be working well.

I am now starting to prepare a proper pull request to add support for faster-whisper alongside whispercpp. However, before I do that, I would like to ask whether this is the right approach and something the project would be interested in.

Perhaps, at least for whisper (in contrast with vosk), it would be preferable to support OpenAI's API so that people can expose their locally running model (with either whisper.cpp or faster-whisper) for use by other tools besides maubot. Furthermore, I find myself forced to add quite a bit of complexity in order to support two different flavors of whisper, which could be avoided by relying on the API.