fix: sử dụng soundfont từ server

2026-07-26 20:25:31 +07:00
parent 2f2cb3a066
commit 40c531087c
4 changed files with 420 additions and 165 deletions
@@ -1,181 +1,206 @@
-# VST3 / SoundFont Engine — Implementation Plan
+# Technical Analysis & Implementation Plan: SF3 + SpessaSynth Client (md/37_SF_CONVERT.md)

-## Current State vs Requirements
+## 1. Assessment Summary

-| Check | Status | Ref |
-|-------|--------|-----|
-| FluidSynth Channel 9 for Drums (bank=128) | ✅ Already implemented. `render_engine.py:64` forces `midi_channel=9` when `is_percussion` or `bank==128`. | R1 verify |
-| Catalog cache invalidation on SF upload | ✅ Already implemented. `plugins.py:93-95` calls `invalidate_catalog_cache()`. | R2 verify |
-| DecentSampler CWD swap for `.wav` samples | ✅ Already implemented. `vst_engine.py:338-341` does `os.chdir(preset_dir)`. | R3 verify |
-| FluidSynth SF2 path uses `static/soundfonts/` hardcoded | ❌ `render_engine.py:186-189` hardcodes `static/soundfonts/<id>.sf2`. Must scan system + upload dirs. | **Bug** |
-| SF2 upload dir (`app/storage/uploads/soundfonts/`) not searched in render_engine | ❌ FluidSynth render branch never looks at user uploads. | **Bug** |
-| `synth_engine` structured metadata not parsed | ❌ Render path uses flat fields only. Spec uses `{type, plugin_id, bank, program}`. | **Gap** |
-| Client preview still oscillator-emulated | ⚠️ Acceptable per spec: "Wasm Module / Preview Synth" — no real SF2 in browser. `soundfontPlayer.js` reads flat fields, not `synth_engine`. | **Gap** |
-| Host asset permissions might deny Docker | ⚠️ Non-root users may get `Permission Denied` on mounted `.vst3`/`.sf2`. | R4 |
-| Missing VST3/SF2 fallback chain | ⚠️ No graceful degraded path if selected instrument is absent. | R5 |
+The spec proposes 2 steps:
+1. **Server**: Convert `.sf2` → `.sf3` (Ogg Vorbis compressed) via `mscore` CLI
+2. **Client**: Replace oscillator emulation with SpessaSynth library for authentic SoundFont playback

---
+### Current State vs Spec

-## Tasks
+| Requirement | Status | Impact |
+|---|---|---|
+| `mscore` in Dockerfile | ❌ Not installed | 200MB+ dependency |
+| `soundfont_converter.py` | ❌ Does not exist | Needs creation |
+| `GET /soundfonts/download/{sf_id}` endpoint | ❌ Missing | Blocks client download |
+| `spessasynth_lib` CDN import in `index.html` | ❌ Not present | Blocks client upgrade |
+| `soundfontStorage.js` (IndexedDB) | ❌ Not created | Needed for caching |
+| `SonicSF.init(audioCtx)` | ❌ Not called anywhere | New integration point |
+| `playNote` signature compat | ⚠️ Spec uses 4 params, codebase uses 6–8 | Must bridge |
+| `applyAITrackInstrument` signature | ⚠️ Spec has `channel` first, code passes it last | Must bridge |

-### Task A — Environment Validation
+### Key Risk: `mscore` Dependency

-1. **Host asset check + permissions**
-   - Verify `ls /home/locpham/daw_assets/{vst3,soundfonts,pianobook}/*` returns files
-   - `chmod -R 755 /home/locpham/daw_assets`
-   - **Files:** host paths only
+`mscore` (MuseScore) pulls in Qt, fontconfig, audio drivers — easily 200-400MB in the container. **Alternative approaches:**

-2. **Docker compose mount verification**
-   - `docker-compose.yml:13-15` maps:
-     - `vst3` → `/opt/daw_engine/vst3`
-     - `soundfonts` → `/opt/daw_engine/soundfonts`
-     - `pianobook` → `/opt/daw_engine/samples/pianobook`
+1. **`fluidsynth` built-in conversion**: `fluidsynth` already installed (`libfluidsynth3`, `pyfluidsynth`). Can convert SF2→SF3 via `--convert` flag or using `fluid_synth_sffd_*` APIs, but the low-level Python bindings don't expose this.

-3. **Runtime Python deps verification**
-   - `docker compose exec web python -c "import pedalboard, fluidsynth; from sf2utils.sf2parse import Sf2File; print('OK')"`
-   - Check server logs for `HAS_PEDALBOARD`, `HAS_PYFLUIDSYNTH` flags in `vst_engine.py`
+2. **`sf2convert`/`sf2pack`**: Smaller tools, but less commonly packaged.

---
+3. **Python + `libsndfile`/`pydub`/`ogg`**: Parse SF2, extract WAV samples, compress to Ogg, rebuild SF3 structure. Complex — would need a SF2 parser and Ogg encoder.

-### Task B — Fix SF2 Path Resolution (Bugfix)
+4. **Python `acousticbrainz-sf2convert`**: Lightweight Python library specifically for SF2↔SF3.

-4. **`render_engine.py` FluidSynth branch** — replace hardcoded `static/soundfonts/` path
-   - **Current** (line 185-189): builds path to `app/static/soundfonts/<sf_id>.sf2` only
-   - **Target**: search in order:
-     1. `UPLOAD_SF_DIR` = `app/storage/uploads/soundfonts/<id>.sf2`
-     2. `SYSTEM_SF_DIR` = `/opt/daw_engine/soundfonts/<id>.sf2`
-     3. Match by `soundfont_id` field in track metadata (not just filename)
-   - Accept both `sf_<id>` and bare `<id>` in `instrument_id`
-   - **Files:** `app/core/render_engine.py:185-225`
+**Recommendation**: Install `mscore` via `apt-get install -y mscore --no-install-recommends` to minimize deps. If the image grows too much (~1.5GB+), fall back to `fluidsynth` command-line conversion (`fluidsynth --convert`).

-5. **Read `soundfont_bank`/`soundfont_program` from track in FluidSynth branch**
-   - Currently read at lines 61-63 (before item loop) — ✅ correct
-   - Ensure `fl.program_select(midi_channel, fid, bank, program)` uses them (line 194) — ✅ correct
-   - **Files:** `app/core/render_engine.py` (verify only)
+### Key Risk: `playNote` API Compatibility

---
-
-### Task C — `synth_engine` Metadata Alignment
-
-6. **`render_engine.py`: parse `synth_engine` object**
-   ```python
-   se = track.get("synth_engine", {})
-   instrument_id = se.get("plugin_id") or track.get("instrument_id", "")
-   soundfont_bank = se.get("soundfont_bank") or track.get("soundfont_bank", 0)
-   soundfont_program = se.get("soundfont_program") or track.get("soundfont_program", 0)
-   instrument_source = se.get("type") or track.get("instrument_source", "soundfont")
-   ```
-   - Apply before the item loop (around line 57-67)
-   - Keep flat fields as fallback for backward compat
-   - **Files:** `app/core/render_engine.py`
-
-7. **`app.jsx`: write `synth_engine` alongside flat fields**
-   - In `setTrackInstrumentWithProgram()` (line 6490): add `synth_engine: {type, plugin_id, soundfont_bank, soundfont_program}`
-   - In `setTrackInstrument()` (line 6540): same
-   - Type logic:
-     - `instrumentId` starts with `sf_` → `type: "soundfont"`
-     - `instrumentId` is VST name → `type: "vst3"`
-     - `null` → `type: "default"`
-   - **Files:** `app/static/js/app.jsx:6490-6561`
-
-8. **`aiGateway.js`: include `synth_engine` in track context**
-   - In `buildAIPromptContext()` (line 218): add `synth_engine` to track objects
-   - **Files:** `app/static/js/services/aiGateway.js:218-245`
-
---
-
-### Task D — Graceful Fallback Chain (R5)
-
-9. **`render_engine.py`: 3-level fallback for missing instruments**
-   - Level 1: Selected VST3/SoundFont
-   - Level 2: Default `GeneralUser_GS.sf2` (or first available `.sf2`)
-   - Level 3: Basic oscillator synth (`render_midi_events_to_audio`)
-   - **Files:** `app/core/render_engine.py:131-230`
-
---
-
-### Task E — Client Playback Enhancement
-
-10. **`soundfontPlayer.js`: read `synth_engine` from track context**
-    - `playNote()` accepts optional `synthEngine` param
-    - Before scheduling: call `controllerChange(ch, 0, bank)` and `programChange(ch, program)`
-    - Use existing oscillator ADSR emulation (no WASM SF2 — scope limit)
-    - **Files:** `app/static/js/services/soundfontPlayer.js:82-204`
-
-11. **`app.jsx`: pass `synth_engine` to `SonicSF.playNote`**
-    - In `schedulePianoRollMidi()`: read `synth_engine` from active track, forward it
-    - **Files:** `app/static/js/app.jsx`
-
---
-
-### Task F — AI Workflow Verification
-
-12. **Catalog API test**
-    - `GET /api/v1/plugins/soundfonts/catalog` returns `{full_catalog, condensed_catalog}`
-    - Condensed ≤ 50 entries
-    - `window.__soundfontCatalog` populated on app load
-
-13. **AI prompt injection test**
-    - `buildCatalogPromptSection()` generates catalog text with bank/program rules
-    - Submit: *"Compose 8 bars of Brass horns and a drum kit"*
-    - Verify returned JSON: Brass has `bank:0, program:56`, Drums has `bank:128, program:0`
-
---
-
-### Task G — End-to-End Test
-
-14. **Manual checklist**
-    - [ ] `docker compose up --build` succeeds
-    - [ ] Synth button shows "Synth" → click → dropdown lists SoundFonts + VSTs
-    - [ ] Select SoundFont → button label updates → instrument presets appear
-    - [ ] Select preset → label reflects exact instrument name
-    - [ ] Draw MIDI notes → Play → oscillator preview (approximate GM sound)
-    - [ ] Export WAV → file plays correct FluidSynth/VST3 instrument
-    - [ ] Select "None (Default Synth)" → oscillator fallback works
-    - [ ] AI generates track with instrument → export plays correct patch
-    - [ ] Upload new `.sf2` → appears in dropdown after refresh
-
---
-
-### Task H — Install Missing Assets (if needed)
-
-15. If `.sf2` absent: copy `GeneralUser_GS.sf2` or `SGM-V2.01.sf2` to `/home/locpham/daw_assets/soundfonts/`
-16. If `.vst3` absent: place `DecentSampler.vst3` in `/home/locpham/daw_assets/vst3/`
-17. If `.dspreset` absent: place Pianobook library in `/home/locpham/daw_assets/pianobook/`
-
---
-
-## Affected Files
-
-| File | Changes |
-|------|---------|
-| `app/core/render_engine.py` | Fix SF2 path resolution (Task B), parse `synth_engine` (Task C), 3-level fallback (Task D) |
-| `app/static/js/app.jsx` | Write `synth_engine` in setTrackInstrument functions (Task C), pass to SoundFontPlayer (Task E) |
-| `app/static/js/services/soundfontPlayer.js` | Accept `synthEngine` param, dispatch CC/program before note (Task E) |
-| `app/static/js/services/aiGateway.js` | Include `synth_engine` in AI track context (Task C) |
-| `app/api/v1/plugins.py` | No changes needed (cache invalidation already exists) |
-| `app/core/vst_engine.py` | No changes needed (CWD swap already exists) |
-| Host `/home/locpham/daw_assets/*` | `chmod 755`, ensure files exist |
-
-## Validation
-
-```bash
-# 1. Build & boot
-docker compose up --build -d
-
-# 2. Verify deps
-docker compose exec web python -c "import pedalboard, fluidsynth; from sf2utils.sf2parse import Sf2File; print('OK')"
-
-# 3. Catalog API
-curl -s http://localhost:8000/api/v1/plugins/soundfonts/catalog | python -m json.tool | head -60
-
-# 4. Render smoke test — create minimal project JSON and POST /api/v1/plugins/render
+The spec's proposed API:
+```javascript
+playNote(pitch, velocity=0.8, durationSec=1.0, channel=0)
 ```

-## Rollback
-
-All changes backward-compatible (flat fields still work if `synth_engine` absent).
-```bash
-git checkout -- app/core/render_engine.py app/static/js/app.jsx \
-  app/static/js/services/soundfontPlayer.js app/static/js/services/aiGateway.js
+Current callers use:
+```javascript
+playNote(pitch, velocity, durationMs, startTime, program, destinationNode, channel?, synthEngine?)
 ```
+
+**Strategy**: Don't replace. Instead, wrap SpessaSynth inside the existing `SonicSF` object. Map:
+- `velocity` (0–1) → MIDI velocity (1–127)
+- `startTime` → if in future, use `setTimeout` for note scheduling (not perfect but adequate for preview)
+- `destinationNode` → SpessaSynth routes to its own internal destination, but can add a gain node stage
+- `program` → internal `_channels[channel].program` state (as now), but SpessaSynth also gets `programChange(ch, program)`
+- `synthEngine` → load the right SF3 and set bank/program on SpessaSynth
+
+**Backward compat**: Keep ALL existing methods. Add SpessaSynth as an optional enhanced engine. If SpessaSynth isn't loaded (CDN fails), fall back to oscillator emulation.
+
+---
+
+## 2. Implementation Plan
+
+### Task 1 — Server: SF2→SF3 Conversion Pipeline
+
+#### 1.1 Docker Dependency
+- `Dockerfile`: Add `mscore` with `--no-install-recommends`
+  ```dockerfile
+  RUN apt-get install -y --no-install-recommends mscore && rm -rf /var/lib/apt/lists/*
+  ```
+- Test: `docker compose build` — measure image size delta
+
+#### 1.2 `app/core/soundfont_converter.py`
+- `SoundFontConverter` class with:
+  - `convert_sf2_to_sf3(sf2_path)` → runs `mscore -o output.sf3 input.sf2`
+  - `batch_convert_all()` → walks `target_dirs`, converts missing `.sf3`
+  - `target_dirs` = `["/opt/daw_engine/soundfonts", "app/storage/uploads/soundfonts"]`
+  - Progress logging, error handling, cache validation (skip if `.sf3` newer than `.sf2`)
+- Integrate with server startup:
+  - In `app/main.py` startup event, call `SoundFontConverter().batch_convert_all()` as background task (non-blocking, don't delay boot)
+
+#### 1.3 API Download Endpoint
+- `app/api/v1/plugins.py`: Add:
+  ```python
+  @router.get("/soundfonts/download/{sf_id}")
+  async def download_soundfont_asset(sf_id: str):
+      # Search system_dir, upload_dir for .sf3 or .sf2
+      # Return FileResponse
+      # Fallback: .sf3 → .sf2 → 404
+  ```
+- Use `_find_sf2_path` from `render_engine.py` logic (case-insensitive, multi-dir)
+- Set `media_type="application/octet-stream"` and proper `Content-Disposition`
+
+#### 1.4 Update `SoundFontInspector` to Also Scan `.sf3`
+- `soundfont_inspector.py:80`: Change `.sf2` filter to `('.sf2', '.sf3')`
+- Ensure the catalog includes `.sf3` files as available instruments
+
+---
+
+### Task 2 — Client: SpessaSynth Integration
+
+#### 2.1 `index.html` — Add SpessaSynth CDN
+```html
+<script type="module">
+  import { Synthesizer } from 'https://cdn.jsdelivr.net/npm/spessasynth_lib@latest/dist/spessasynth_lib.js';
+  window.SpessaSynthClass = Synthesizer;
+</script>
+```
+- Add BEFORE `soundfontPlayer.js` so the class is available when the player initializes
+
+#### 2.2 `app/static/js/services/soundfontStorage.js` — IndexedDB Cache
+- `SoundFontStorage` class with `openDB()`, `getBuffer(sfId)`, `saveBuffer(sfId, arrayBuffer)`
+- Uses `indexedDB` with DB name `"DAW_SoundFont_Cache"`, store name `"sf3_buffers"`
+- Export singleton `sfStorage`
+
+#### 2.3 `app/static/js/services/soundfontPlayer.js` — Dual-Mode Rewrite
+
+**Architecture**: Keep the existing `window.SonicSF` as the public API. Internally use SpessaSynth when available, fall back to oscillator when not.
+
+**Changes to existing methods:**
+
+| Method | Change |
+|---|---|
+| `init(audioCtx)` | NEW — creates SpessaSynth instance, triggers default SF load |
+| `loadSoundFont(sfId)` | REWORK — try IndexedDB → fetch `/api/v1/plugins/soundfonts/download/{sfId}` → load into SpessaSynth |
+| `playNote(...)` | ADD SpessaSynth path: if initialized, delegate to `synthInstance.noteOn/noteOff`; else use oscillator fallback |
+| `applyAITrackInstrument(bank, program, synthEngine?)` | ADD SpessaSynth path: call `controllerChange` + `programChange` on SpessaSynth |
+| `stopAll()` | ADD SpessaSynth: `allNotesOff(channel)` or `programReset()` |
+| `controllerChange(ch, cc, val)` | ADD SpessaSynth: delegate if initialized |
+| `programChange(ch, prog)` | ADD SpessaSynth: delegate if initialized |
+
+**Backward compat guarantee:**
+- All existing callers continue to work unchanged
+- 6-arg `playNote(pitch, vel, durMs, startTime, program, destNode)` → SpessaSynth ignores `startTime` (schedules immediately) and `destNode` (uses internal routing)
+- 8-arg `playNote(..., channel, synthEngine)` → SpessaSynth uses `synthEngine.soundfont_id` for SF loading, `synthEngine.soundfont_program` for program selection
+- If SpessaSynth not loaded/CDN fails → transparent fallback to existing oscillator code
+
+#### 2.4 `app.jsx` — Integration Points
+
+- **Startup** (in a useEffect or the existing audio context initialization):
+  ```javascript
+  if (window.SonicSF && window.SonicSF.init) {
+    SonicSF.init(getAudioContext());
+  }
+  ```
+- **Synth selection**: After `setTrackInstrumentWithProgram` / `setTrackInstrument`, trigger SF load:
+  ```javascript
+  if (synthEngine && synthEngine.type === 'soundfont') {
+    SonicSF.loadSoundFont(synthEngine.soundfont_id);
+  }
+  ```
+- **AI track creation**: Already calls `applyAITrackInstrument` — SpessaSynth path handles it
+
+---
+
+### Task 3 — Backward Compatibility & Migration
+
+#### 3.1 Fallback Behavior
+- If `window.SpessaSynthClass` is undefined (CDN blocked, offline): fall back to existing oscillator code
+- If `.sf3` download fails: fall back to oscillator
+- If SpessaSynth throws: catch error, log warning, fall back to oscillator
+
+#### 3.2 Testing Matrix
+
+| Scenario | Expected |
+|---|---|
+| SpessaSynth loaded + SF3 cached | Authentic playback, 0ms load |
+| SpessaSynth loaded + SF3 needs download | Authentic playback after 1-2s load |
+| CDN blocked (offline) | Transparent oscillator fallback |
+| SF3 not available on server | Transparent oscillator fallback |
+| MIDI keyboard + armed track | Authentic or oscillator based on availability |
+
+---
+
+## 3. Files Affected
+
+| File | Change |
+|---|---|
+| `Dockerfile` | Add `mscore --no-install-recommends` |
+| `app/core/soundfont_converter.py` | **NEW** — SF2→SF3 convert + batch scan |
+| `app/api/v1/plugins.py` | Add `GET /soundfonts/download/{sf_id}` |
+| `app/core/soundfont_inspector.py` | Accept `.sf3` in scan filter |
+| `app/main.py` | Add startup background conversion |
+| `app/templates/index.html` | Add SpessaSynth CDN `<script type="module">` |
+| `app/static/js/services/soundfontStorage.js` | **NEW** — IndexedDB cache |
+| `app/static/js/services/soundfontPlayer.js` | Dual-mode rewrite (SpessaSynth + oscillator fallback) |
+| `app/static/js/app.jsx` | Add init + SF load calls |
+
+## 4. Validation
+
+```bash
+# Server-side
+docker compose build  # verify mscore installs
+docker compose exec web python -c "from app.core.soundfont_converter import SoundFontConverter; print('OK')"
+curl -s http://localhost:8000/api/v1/plugins/soundfonts/download/sgm_v2.01 | head -c 4 | file -
+docker compose exec web ls -la /opt/daw_engine/soundfonts/*.sf3  # verify conversion
+
+# Client-side
+# Open browser → DevTools → check SpessaSynth loaded (window.SpessaSynthClass)
+# Select SoundFont → check network tab for .sf3 download
+# Play MIDI notes → hear authentic instrument (not oscillator)
+```
+
+## 5. Open Questions
+
+1. **`mscore` image size**: Measure actual delta. If >300MB, consider `fluidsynth --convert` alternative.
+2. **CDN reliability**: SpessaSynth loaded from jsdelivr CDN — consider bundling or NPM install as fallback.
+3. **Startup delay**: `batch_convert_all()` could take minutes for large SF2s. Run as background Celery task, not inline startup.
+4. **SGM_v2.01.sf2 (529MB)**: Conversion time for this file. Need to handle gracefully (streaming, timeout).