TUI voice mode: Python crash on 2nd transcription (cTranslate2/OpenBLAS threading conflict)

[Montbra]

Bug Description

TUI voice mode crashes Python process on the 2nd transcription attempt. First transcription works, 2nd always crashes with SIGABRT.

CLI ancient (without --tui) works correctly — no crashes, Ctrl+B stops and transcribes, auto-stop by silence works, TTS works.

Steps to Reproduce

1. Start Hermes with TUI: hermes --tui
2. Enable voice mode: /voice on
3. Press Ctrl+B to start recording, speak, press Ctrl+B to stop — 1st transcription works ✓
4. Press Ctrl+B again to start new recording, speak briefly, press Ctrl+B — Python crashes (SIGABRT)

Environment

• OS: Linux (openSUSE Tumbleweed)
• Hermes: latest from NousResearch/hermes-agent
• Python: 3.11
• Audio: sounddevice + PortAudio + PipeWire
• STT: faster-whisper (local, GPU)

Crash Stack Trace

frame.abi3.so + CFFI + libportaudio + scipy_openblas
__pyx_pf_2av_5frame_5Frame_2__dealloc__
PyThread_acquire_lock_timed + _queue.cpython-311
epoll_wait + libspa-support.so + libpipewire-0.3

Multiple threads: libgomp, blas_thread_server, ctranslate2 Worker, libpipewire.

Root Cause

TUI uses hermes_cli/voice.py with start_continuous(). The 2nd transcription crashes — likely accumulated state in cTranslate2/OpenBLAS thread pool. CLI ancient avoids this because each transcription runs in a fresh thread that terminates.

Workaround

Use CLI ancient (without --tui). Voice mode works correctly there.

Severity

Medium — TUI voice unusable. CLI ancient is working alternative.

[dhruvac29]

Hey @Montbra, great report! Would you be up for submitting a PR to fix this?

[Montbra]

Hi @dhruvac29, I will dig deeper and return later

[dhruvac29]

Sounds good, take your time! If you need any pointers on the codebase or want to talk through the fix, feel free to ping here.

[Montbra]

Resolution

The TUI voice mode crash on 2nd transcription has been fixed. The root causes were:

Fix 1: ui-tui/src/app/useInputHandlers.ts

Added explicit session_id to the RPC call:

gateway.rpc('voice.record', { action, session_id: getUiState().sid })

Previously the payload only had { action } without a session ID.

Fix 2: tui_gateway/server.py

Three critical fixes in the voice.record handler:

1. Safe dict cast: Added if not isinstance(voice_cfg, dict): voice_cfg = {} to prevent silent AttributeError crashes
2. auto_restart=False: Changed start_continuous call to pass auto_restart=False — prevents infinite VAD loop after first silence-triggered stop
3. force_transcribe=True: Changed stop action to call stop_continuous(force_transcribe=True) instead of stop_continuous()

Fix 3: hermes_cli/voice.py

1. auto_restart parameter in start_continuous: The silence callback now checks this flag and cleanly stops instead of restarting the microphone
2. force_transcribe in stop_continuous: When True, creates a thread that calls rec.stop() (preserves audio) instead of rec.cancel() (discards audio), then transcribes and emits the result

Key insight

The original stop_continuous() always called rec.cancel() which discards buffered audio. The 2nd transcription crash was caused by audio being discarded, not by cTranslate2/threading issues as originally suspected.

Status: RESOLVED