Test live audio in Redpanda-vc
Choose desktop or browser
The Windows desktop realtime launcher uses the devices exposed to that local process. The WebUI realtime tab captures microphone input through the browser and passes audio to the selected inference device. Use one path for your first test and leave the other closed to make feedback and competing jobs easier to diagnose.
- Load a compatible RVC model, select the expected embedder where exposed, and start at neutral pitch with the default buffer.
- Select the physical microphone. In browser mode, grant mic permission to the local page and check the actual source in browser permissions.
- Connect headphones and play the monitor. Say one short sentence containing both quiet and loud words, followed by a pause. Check voice identity, volume, and dropouts.
- Stop the stream before switching engines or choosing a different inference device. Run another test after reconnecting.
Select the compute device
In WebUI builds with Inference Device, Automatic assigns a new stream to an available device. Choose a specific GPU when comparing repeatable results or leaving a GPU free for another workload. A stream stays on its assigned device. Multiple GPUs can handle separate streams; changing the selector is not a way to combine their capacity into one stream. Beatrice uses a separate CPU path in the documented implementation.
Route the signal after a clean test
If your playback path can target a virtual cable, choose the cable's playback endpoint there. In the receiving app, select the companion recording endpoint as the microphone. Keep the physical mic as the converter input and output from the call on headphones. In browser mode the playback destination can depend on browser and OS output settings. Verify the receiver with a local recording before judging the live stream in a call.
The browser stream attempts WebSocket transport and can fall back to HTTP streaming when needed. Establish a clean local result before investigating remote connection quality.