- Remove the 'instruct' style field: Qwen3-TTS Base generate_voice_clone
silently ignores it, so the UI option did nothing.
- Serialize generations with a lock so concurrent requests queue instead
of doubling RAM/VRAM on one shared model.
- Validate empty text/ref_text ourselves so the API returns the Korean
400 messages instead of FastAPI's generic 422.
- UI: explain that mic recording needs https/localhost instead of a raw
TypeError on plain-http LAN access.
- README: document mic/https limit, queueing, and measured CPU speed.
Verified on CPU (Python 3.12, torch 2.14 CPU, 0.6B Base): wav and mp3
references, language ko/auto, two concurrent requests -> all 200 and
Whisper transcripts match the requested text.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
a34f16123f