What's New
Unified Participant Audio Capture on Linux
-
Linux now uses the same audio capture method as Windows and macOS for participant (tab/system) audio — no more manual PulseAudio/PipeWire virtual device setup. The desktop app handles participant audio loopback natively on all three platforms. (#192)
-
Participant audio restores automatically on app restart — previously Linux users could lose participant audio after restarting the app.
-
TTS virtual devices (used for injecting translated speech into meetings) remain unchanged.
Local Inference — Memory Usage Visibility
- Estimated VRAM/RAM usage is now shown in the model info panel when using the Local Inference provider. The estimate reflects the combined memory footprint of your selected main and participant models, split by device type (GPU vs CPU). Cloud-based models like Edge TTS are excluded from the estimate.
Local Inference — GPU Out-of-Memory Protection
-
User-friendly error when GPU runs out of memory — if WebGPU cannot allocate enough VRAM during model loading, Sokuji now catches the error and shows a translated message advising you to switch to a smaller model, instead of silently failing.
-
Pre-flight memory budget check — before loading participant models, Sokuji estimates whether the GPU/system has enough memory available and warns you early if the model is too large.
Compatibility
- Chrome / Edge 116+ (extension)
- Desktop app: Windows / macOS / Linux