How to give raw audio file to stream instead of microphone?

[SmartManoj]

I am trying to live captioning a call.

[ggerganov]

One way would be to make a loopback audio capture device on your system and select it via the -c argument of stream.
Another approach would be something like the twitch.sh stream tool.

[tomheno]

Hi, @ggerganov
In the case of twitch.sh, main is called numerous of time, meaning the model is loaded at each iteration, right ?

whisper.cpp/examples/twitch.sh

Line 103 in 4ef3398

./main -t $threads -m ./models/ggml-$model.bin -f /tmp/whisper-live.wav --no-timestamps -otxt 2> /tmp/whispererr | tail -n 1

Would there be a way to avoid that, and keep it warm in the memory, as it's a continuous process ?
Thanks a lot for the job done on this library !

[ggerganov]

@tomheno
Maybe the idea from #182 (reply in thread) can be utilised for this:

• start main and configure it to monitor a folder for new .wav files
• when there is a new .wav file - transcribe it
• start a separate process that reads twitch stream audio and chunks it into .wav files -- the files are written to the folder monitored by main

[pbengert]

Here is what you can do to use audio output instead of the microphone:
Get VB-Cable from https://vb-audio.com/Cable/
It will install a virtual audio cable. With that you can reroute sound from sound output to input. Choose VB Cable as your "speaker" and on the other side as your "microphone". That way you can use any sound output with stream. That works nice with e.g. YouTube.
When you start "stream" you can find the capture devices at the beginning of the init output. Mine looks like that:
init: - Capture device #2: 'VB-Cable'
So you append: -c 2to the command to use this a your sound source.