Improvements
- Local transcription no longer drops speech. ASR decoding now runs outside the voice-activity detector, so audio keeps being captured while a segment is still being transcribed. On a 70 s continuous Japanese stream, Whisper Small previously captured only 32 s — two clips vanished entirely — and now captures all of it with no errors (#497, contributed by atirna).
- Conversation display filter: a new mode hides one side of the conversation entirely, one click from the default now goes straight to translation-only, and the toolbar buttons draw the lines that are showing rather than only signalling that a filter is on (#512).
- The conversation toolbar is easier to read, its icons resting at a 4.8:1 contrast ratio instead of 2.2:1 (#512).
- Sokuji no longer sends an internal audio-quality event every 30 seconds; every field it carried was a placeholder.
Bug fixes
- A machine with a working GPU is no longer misdetected as having none. While the GPU process was still starting, the first WebGPU query could answer "no adapter", and Sokuji cached that for the rest of the session — hiding models you had already downloaded, refusing f16 model variants, and showing a false "no GPU acceleration" notice. Measured on Windows with an RTX 4070 SUPER (#515).
- Granite Speech no longer crashes on long segments. Any segment reaching the 20-second speech cap aborted during transcription; in an 8-clip test every such segment failed before and none do now (#502, fixes #500).
- f16 model failures now name the model and say what to do. Loading an f16 variant on a GPU without that feature produced "Program Transpose requires f16 but the device does not support it", naming nothing; the check now runs against the adapter the model actually loads on, across all ten WebGPU workers (#506, #514, fixes #504).
Note
- Startup diagnostics now report the GPU adapter and OS version, so a bug report can be answered without first asking what hardware it came from (#505).
Full Changelog: v0.40.2...v0.40.3