[raphaelm22]

Draft: I'm running this on my real meetings before marking it ready. Design feedback is welcome now.

Summary

Implements #131. A third Compute Device, Intel GPU (OpenVINO), runs Whisper on Intel integrated GPUs through OpenVINO GenAI's WhisperPipeline. faster-whisper (CTranslate2) only supports the CPU and NVIDIA.

On a Core Ultra 7 265U (Intel Graphics iGPU), the test was a real 7.4 min pt-BR meeting with large-v3-turbo:

	Transcription
CPU, faster-whisper (today)	7.7 min
Intel GPU, this PR (in the app's worker)	40–49 s
first use: +34 s model download, +~40 s GPU compile (cached afterwards)	83 s

On a 60 s excerpt the text matched faster-whisper's word for word, and segment start/end times are the same.

How it works

• app/transcription/openvino_whisper.py (no Qt; OpenVINO and faster-whisper imported inside functions):
  • Intel's int8 builds OpenVINO/whisper-<size>-int8-ov, one per model size in Settings, downloaded to the normal HF cache.
  • A resident pipeline cache like _MODEL_CACHE. The compiled model is cached in ~/.talktrack/openvino_cache.
• VAD is ours to do. WhisperPipeline has none, and Whisper invents text over silence: 20 s of silence came back as "E aí". faster-whisper's own Silero VAD picks the speech, only speech is decoded, and a TimeMap maps times back to the recording.
• Batches of ≤ 60 s of speech. generate() blocks, and its streamer is not available with timestamps on long audio, so batches are the only progress and cancel points.
  • Meeting speech rarely pauses for the VAD's 2 s; one region ran 264 s. So the VAD also gets max_speech_duration_s=60, which splits long speech at a short pause, as the batched pipeline does at 30 s.
  • This brought Cancel from ~20 s to 5 s at the same total time. Compared with 120 s batches, it also stopped some phrases from being dropped.
• Language is detected on the first batch, then pinned, as faster-whisper detects once per recording.
• TranscriptionWorker._run_openvino: any OpenVINO failure falls back to the CPU, at load (package missing, no GPU) or mid-run (driver hang); the pipeline is evicted and the user sees the reason in the progress line. Unreadable audio still errors, since the CPU would fail too.
  • This is broader than the diarizer's OOM-only retry, on purpose. These failures are environmental, and the transcript matters more than surfacing them.
• Settings: new device entry with inline status (ready / not installed / no Intel GPU). openvino-genai (~90 MB download) is installed on demand through the existing package_installer, as the AI SDKs are.
• System Status:
  • An Intel GPU (OpenVINO) row appears only when that device is selected, since checks may now return None.
  • The Whisper model check looks for the OpenVINO build.
  • GPU Acceleration no longer tells Intel GPU users to switch to CPU.
• Packaging:
  • openvino extra in pyproject.toml, plus a commented line in requirements.txt. One test keeps the three pip specs equal.
  • uv.lock adds only the four new packages. As in fix: cap PyAV below 19 so faster-whisper can decode audio (#127) #130, I hand-applied them over master's lock to avoid uv's marker reordering, and uv lock --check passes on uv 0.12.7 and 0.13.0.
  • The module is added to the CI import smoke.
• Diarization stays on the CPU (_resolve_device("intel_gpu") → cpu). OpenVINO ships TBB rather than another libiomp5md.dll, so App deadlocks at zero CPU when diarization follows transcription (duplicate libiomp5md.dll) #70's duplicate-OpenMP deadlock doesn't apply. pyannote ran normally in the same process after OpenVINO.

Worth reviewing

• Config.set() saves immediately. The OpenVINO install prompt therefore runs before any config.set() in _save_and_close, so declining leaves settings.json untouched. The existing AI-SDK prompt sits after the sets, so its return # User cancelled install, don't save does not actually prevent the save. I left that alone since it's out of scope here.
• confidence stays 0.0 for these segments. OpenVINO reports no per-segment log-probability, and nothing reads the field today.
• No NPU option. On this machine's NPU, large-v3-turbo hangs the device (ZE_RESULT_ERROR_DEVICE_LOST, 2/2 runs), and only small ran (35 s for the same meeting).

Tests

• New tests (49):
  • tests/test_openvino_whisper.py: model repos, batching, time mapping, the transcribe loop against a fake pipeline, language pinning, GPU detection, the pipeline cache and download, the pip-spec sync and the Settings prompt order.
  • Worker tests in test_transcriber.py: GPU path, fallback at load and mid-run, cancel, unreadable audio.
  • System Status tests in test_dependency_checker.py.
• Full suite: 457 passed.
• End to end: inside an offscreen QApplication with torch loaded first, as main.py does, I ran the real TranscriptionWorker on the meeting:
  • first use (download and compile) and a warm run
  • cancel after 8 s
  • CPU fallback with openvino_genai blocked
  • DiarizationWorker afterwards in the same process
• Settings: exercised with real widgets offscreen, with the config's save disabled so nothing was written.

Closes #131

🤖 Generated with Claude Code

[raphaelm22]

Ready for review. Used daily since 15/10 on a Core Ultra 7 265U (Intel GPU selected, live transcription on): 7 real Teams meetings, 11 to 115 min, pt-BR, 4 to 6 speakers. No fallbacks or errors in the log. Whisper large-v3-turbo on the iGPU throughout, no CPU fallback needed.