How to transcribe from audio output device?

[playgithub]

transcribe from audio output device on linux,
output on screen as stream,
ignore audio input devices (microphone)

[bogdanr]

Use the monitor source of your output device. PulseAudio and PipeWire make a monitor source for each sink. A monitor source records the audio that the sink plays.

1. Build the stream tool with SDL2.

sudo apt-get install libsdl2-dev
cmake -B build -DWHISPER_SDL2=ON
cmake --build build --config Release

2. Find the monitor source.

pactl list short sources | grep monitor

The name ends with .monitor. Example:

alsa_output.pci-0000_00_1f.3.analog-stereo.monitor

3. Start the transcription.

Set PULSE_SOURCE for the command. PulseAudio connects this client to the monitor source. The tool does not use the microphone.

PULSE_SOURCE=alsa_output.pci-0000_00_1f.3.analog-stereo.monitor \
  ./build/bin/whisper-stream -m ./models/ggml-base.en.bin -t 6 --step 0 --length 30000 -vth 0.6

The text goes to the screen as a stream.

Notes

• --step 0 starts the sliding window mode with VAD. The tool transcribes after it detects a pause. This gives better text than a fixed interval.

• SDL2 usually does not show monitor sources in its capture device list. Do not use the device index option. The PULSE_SOURCE variable is more reliable.

• pipewire-pulse also accepts PULSE_SOURCE. Use the same commands on PipeWire.

• To find the monitor source of the sink that is currently active, use this command:

  pactl get-default-sink | sed 's/$/.monitor/'

I tested the routing on PipeWire. PULSE_SOURCE moved the capture stream to the monitor source, and not to the microphone:

$ PULSE_SOURCE=...HiFi__Speaker__sink.monitor parec --raw > /tmp/loop.raw
$ pactl list source-outputs
Source Output #78853
        Source: 54                       # 54 = the monitor source
        application.name = "parec"