- 'Why Soundgen?' section with AI feedback loop diagram
- Training pipeline section: 5-step guide (connect MCP, AI generates, rate, AI improves, export)
- All 5 MCP tools documented in table
- Updated architecture table with soundgen-feedback crate
- SoundSpec example now includes vibrato field
Vibrato (LFO frequency modulation):
- New VibratoSpec in SoundSpec: rate (Hz), depth (cents), delay (s)
- ChannelRenderer applies vibrato to frequency each tick
- Supported on Pulse and Triangle channels
- GUI: vibrato controls (enable, rate, depth) in channel panel
- Enables bird chirps, sirens, wobbles
Feedback loop fixes:
- generate_batch now accepts full SoundSpec array (not just names)
so each sound is unique
- render_sound returns reference examples from feedback DB
- Fixed GUI bug: clicking stars no longer erases feedback text
- Feedback text auto-saves on Enter / focus loss
- Replaced emoji buttons (✓/🗑) with text (Save/Del)
- Green saved indicator when feedback exists in DB
99 tests passing, 0 warnings
New crate: soundgen-feedback
- FeedbackDB with SQLite (rusqlite, bundled)
- CRUD: add, update_rating, get_all, get_unrated, top_rated, delete
- search_similar: keyword-based similarity for few-shot reference
- export_jsonl: export rated examples for fine-tuning
- 9 tests
MCP server: 5 tools (was 3)
- generate_batch: generate multiple sounds by name, store in DB
- get_reference_sounds: search DB for highly-rated similar sounds
- render_sound: now returns reference examples from DB in response
- --db flag to specify feedback database path
GUI: Training tab
- Open/create feedback DB (file dialog)
- Generate batch dialog (enter names, output dir)
- Sound list with star rating (1-5), feedback text, play button
- Export dataset as JSONL
- Stats display (total/rated/avg rating)
OpenCode config updated with --db feedback.db path
Emil