Confirm working on Darwin

[aholten]

Had a potential issue with Python 3.10 where a bugfix was needed. Might need to have python 3.11 as floor

[aholten]

Research notes on macOS crash symptoms

Recalled symptom: stttui crashes on macOS at the moment recording stops and transcription begins.

Thread-flow context

In both TUI and CLI modes, model.transcribe() ends up being the first entry into CTranslate2 and it runs on a non-main thread:

• TUI — speech_to_text.py:459-460: stop_recording() on Textual main thread, then run_worker(_transcribe_worker, thread=True) → engine.transcribe() on a worker thread.
• CLI — _create_hotkey_listener at speech_to_text.py:204 spawns a threading.Thread for toggle_recording, which calls stop_recording() + transcribe() on that thread.

Top hypothesis: OpenMP duplicate (CTranslate2 thread-pool init)

CTranslate2 ships its own libomp.dylib; numpy/scipy may already have linked a different one. On macOS the runtime aborts with "OMP: Error #15: multiple copies of OpenMP runtime have been linked..." the first time CTranslate2 spins up its thread pool — which is deferred to the first transcribe() call. Matches the symptom exactly.

Related upstream reports:

• Transcribe method kills jupyter notebook kernel SYSTRAN/faster-whisper#289
• Transcribe crash jupyter notebook kernels SYSTRAN/faster-whisper#820
• https://discuss.huggingface.co/t/segfault-during-pytorch-transformers-inference-on-apple-silicon-m4-libomp-dylib-crash-on-layernorm/160930

Diagnostic: run KMP_DUPLICATE_LIB_OK=TRUE stttui on a failing Mac. If the crash goes away → confirmed.

Secondary hypothesis: sounddevice stream.close() race

PortAudio's macOS backend has documented intermittent segfaults when close() is called while callbacks are still in-flight. stop_recording() does stop(); close() back-to-back with no wait. Reference: spatialaudio/python-sounddevice#78

Diagnostic: abort() instead of stop(), or a short wait between stop() and close().

Proposed fixes (pending repro)

1. Warmup the model on the loader thread. Add a no-op self.model.transcribe(<silence.wav>) at the end of load_model() so CTranslate2's thread pool initializes on the main/loader thread before any worker-thread transcribe. Cheap, no deps, addresses the OpenMP hypothesis.
2. Guard the audio stream teardown. Swap stream.stop() + stream.close() for stream.abort() + stream.close() (and/or a brief wait). Addresses the PortAudio hypothesis.
3. Document KMP_DUPLICATE_LIB_OK=TRUE as an escape hatch in the README's macOS section for users who still hit it.

Python-version note

The existing body mentions a Python 3.10 bugfix; maintainer also recalls a "weird 3.12 thing on Mac" but isn't certain which version — may have conflated the pynput 3.12 reports (pynput #577) with this transcribe crash. Worth retesting across 3.10 / 3.11 / 3.12 / 3.13 once the crash is reproduced.