Highlights
Long sessions no longer degrade
- Fixed audio throughput collapsing — and then transcription stopping altogether — partway through a long session on OpenAI-compatible providers. With turn detection on, which is the default, every captured chunk was appended to an input buffer that was never cleared, and the merge that rebuilt it ran element by element on the main thread. By nine minutes the buffer was around 25 MB and each append cost roughly 58 ms out of a 170 ms budget, so the main thread was starving the audio worklet feeding it and the WebSocket answering it at the same time. Nothing reported an error; the session simply got worse the longer you ran it. (#407, closes #406)
- A failed realtime send is now handled instead of thrown, so one bad send no longer takes the session down with it. (#407)
macOS virtual audio
- Fixed other applications hearing silence from
SokujiVirtualAudiowhile every diagnostic said the device was fine. macOS stores each device's volume itself and pushes it back onto the driver as the device registers — so the stored value survives reinstalling Sokuji — and upgrading to macOS 26 was measured leaving that value at half. Half is not half volume here: the control is logarithmic across a 64 dB range, so it is −32 dB, about 2.5% of the amplitude. Audio still flowed and no layer reported an error, which is why the device stayed visible everywhere it should be while the level meter never moved and the far end heard nothing. Sokuji now puts the device back to unity gain, and unmutes it, when it starts. (#403, reported in #381)
Soniox
- Soniox text-to-speech now runs on
tts-rt-v2, ahead of v1 being withdrawn upstream. (#405)
Full Changelog: v0.36.1...v0.36.2