Setting to enable live preview for Qwen3-ASR (delayed is fine) on fast Macs

[nikolai-vysotskyi]

Context. Mixed Russian/English dictation on a MacBook Pro M5 Max, v1.6.13-beta.1. In a side-by-side test with the same ~35 s text, Qwen3-ASR 1.7B was clearly more faithful than Parakeet v3: it kept numbers and words as spoken, while Parakeet silently changed "350" to "300" and swapped verbs. Qwen's final pass is already fast (~1.4 s for ~35 s of audio, transcribe.cpp + Metal, Q8_0), but supportsStreaming = false is hardcoded, so there is no live text while speaking.

Request: show Qwen3-ASR live text, even with a delay. It doesn't need to be true real-time; text appearing a second or two behind is fine. The final text can still come from the full-buffer pass on stop, so accuracy stays the same.

Make it a user setting instead of a hardcoded limit. Something like "Live preview for heavy models (Qwen / Whisper Large)" in Settings, off by default, maybe with a note that it needs a fast Mac. Hardware moves fast: on M5 Pro/Max/Ultra, and M6-class chips next, Qwen is already fast enough for a delayed preview, and the hardcoded "too slow for real-time" flag will only get more outdated. Users with fast machines could opt in without waiting for the app to whitelist their chip.

Efficiency ideas: expose the CoreML/ANE Qwen path (currently behind qwenPreviewEnabled) for lower power on battery; consider MLX as Fluid Intelligence already does. Keep Q8 for non-English, since 4-bit quants noticeably hurt Russian accuracy in public benchmarks.

Happy to test builds on M5 Max with Russian/English dictation.

AI disclosure: drafted with Claude; the test results above are from my own dictations.