- Use 1.7B model (0.6B had tensor mismatch with cached prompts)
- Speak endpoint uses ref_audio directly (not cached pkl) as fallback
- Cache voice clone prompts in memory on startup
- Add SpeakableText component: 🔊 icon on each p and li element
- Remove old TTSReader sequential approach
- Add global exception handler to TTS server
- Fix profile localStorage caching
- inference_mode + bf16 optimization
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Notes:
- notes table with TEXT/AUDIO types, category support
- Audio upload → OpenRouter Gemini STT → OCI GenAI polish/summary
- Raw STT saved separately in raw_content column
- Polish/summary button for manual re-processing
- Async processing with real-time polling
Voice Clone TTS:
- Qwen3-TTS 1.7B model on A10 GPU via FastAPI server
- Voice profile registration (record/upload → save embedding)
- Profile-based TTS generation API
- TTS web page with recording, profile management, generation
Auth fixes:
- Store both access + refresh tokens in localStorage
- Initialize state from localStorage synchronously (no flash)
- Request interceptor reads token from localStorage every request
- Refresh via body (not just cookie)
Other fixes:
- maxTokens 4096 → 65536 (OCI GenAI Gemini supports up to 65536)
- Fix broken Korean chars in source files
- OpenRouter config for STT
- ffmpeg installed for audio conversion
- Ollama + Gemma 4 E4B installed (STT fallback)
- nginx proxy for TTS server (/api/tts/)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
