File size: 12,121 Bytes
c27ae8d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75d3906
c27ae8d
75d3906
c27ae8d
75d3906
 
 
 
 
 
 
 
 
 
c27ae8d
75d3906
 
 
 
c27ae8d
 
 
 
 
75d3906
 
c27ae8d
75d3906
c27ae8d
75d3906
c27ae8d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
# Technology Stack & Decisions

## Overview

This document details the technology choices for Rescored, including alternatives considered and trade-offs that informed each decision.

## Frontend Technologies

### UI Framework: React

**Chosen**: React 18+

**Why**:
- Largest ecosystem for music-related JavaScript libraries
- VexFlow and Tone.js have good React integration patterns
- Component model fits notation editing (each measure/staff as component)
- Excellent dev tooling (React DevTools, Fast Refresh)
- Familiarity and hiring pool

**Alternatives Considered**:

| Option | Pros | Cons | Why Not Chosen |
|--------|------|------|----------------|
| Vue 3 | Simpler API, lighter weight | Smaller ecosystem for music libraries | Less community support for music notation |
| Svelte | Excellent performance, less boilerplate | Immature ecosystem | Risk for complex audio/notation needs |
| Vanilla JS | Full control, no framework overhead | Much more code to manage state | Notation editing is complex, need good state management |

**Decision**: React's ecosystem and component model outweigh its learning curve.

---

### Notation Rendering: VexFlow

**Chosen**: VexFlow 4.x

**Why**:
- Pure JavaScript, runs entirely in browser
- Programmatic API for rendering notation (good for editing)
- Generates clean SVG that we can attach event listeners to
- Active maintenance, good documentation
- Used in production by Flat.io, Soundslice

**Alternatives Considered**:

| Option | Pros | Cons | Why Not Chosen |
|--------|------|------|----------------|
| OpenSheetMusicDisplay (OSMD) | Better MusicXML support, prettier output | Harder to build editing on top, heavier bundle | Optimized for display, not editing |
| music21.js | Pythonic API, good theory support | Limited rendering, not designed for web | Better as backend tool |
| abcjs | Lightweight, simple syntax | ABC notation less standard than MusicXML | MusicXML is industry standard |
| Custom renderer | Full control | Months of work to match VexFlow quality | Not worth reinventing wheel |

**Decision**: VexFlow strikes the best balance between rendering quality and edit-ability.

---

### Audio Playback: Tone.js

**Chosen**: Tone.js 14+

**Why**:
- High-level abstractions over Web Audio API
- Built-in scheduling for precise timing
- Multiple synthesis methods (samples, FM, AM)
- Transport controls (play, pause, seek, loop)
- MIDI playback support via `Tone.Sampler`

**Alternatives Considered**:

| Option | Pros | Cons | Why Not Chosen |
|--------|------|------|----------------|
| Web Audio API (raw) | Maximum control, no dependencies | Requires lots of boilerplate | Too low-level for quick MVP |
| Howler.js | Simple API, good for sound effects | Not designed for music, no MIDI | No timing control for notation sync |
| MIDIjs | Simple MIDI playback | Limited synthesis, GM soundfonts | Lower quality sound than Tone.js samplers |
| SoundFont2.js | Authentic GM sounds | Large file sizes, older API | Tone.js can load SoundFonts if needed |

**Decision**: Tone.js provides the right abstraction level for MIDI playback with good sound quality.

---

### State Management: Zustand

**Chosen**: Zustand (tentative)

**Why**:
- Minimal boilerplate compared to Redux
- Works well with React hooks
- Good for global state (notation data, playback state)
- Small bundle size (~1KB)

**Alternatives Considered**:

| Option | Pros | Cons | Why Not Chosen |
|--------|------|------|----------------|
| Redux Toolkit | Battle-tested, great DevTools | More boilerplate, steeper learning curve | Overkill for MVP |
| React Context | Built-in, no deps | Performance issues with frequent updates | Notation editing has lots of updates |
| Jotai/Recoil | Atomic state, very modern | Newer, smaller ecosystem | Zustand more proven |
| Local state only | Simplest | Hard to share state across components | Need global notation state |

**Decision**: Zustand for MVP, can migrate to Redux if needed later.

---

## Backend Technologies

### API Framework: FastAPI

**Chosen**: FastAPI (Python 3.11+)

**Why**:
- Async Python (critical for WebSocket connections)
- Auto-generated OpenAPI docs (Swagger UI)
- Native WebSocket support
- Type hints for better code quality
- Integrates well with Python ML libraries (Demucs, basic-pitch)
- Excellent performance (on par with Node.js)

**Alternatives Considered**:

| Option | Pros | Cons | Why Not Chosen |
|--------|------|------|----------------|
| Node.js (Express) | Async by default, JavaScript everywhere | Worse ML library support | ML models are Python-first |
| Flask | Simple, well-known | No async support, manual WebSocket setup | FastAPI is modern Flask |
| Django | Full-featured, admin panel | Heavy, slower, less async support | Overkill for API-only service |
| Go (Gin/Fiber) | Excellent performance | Weaker ML ecosystem, FFI overhead | Python has better audio/ML tools |

**Decision**: FastAPI combines async support with Python's ML ecosystem.

---

### Task Queue: Celery + Redis

**Chosen**: Celery 5.x with Redis as broker

**Why**:
- Industry standard for async Python tasks
- Reliable, battle-tested in production
- Priority queues (transcription vs. export jobs)
- Automatic retries and error handling
- Redis is fast, simple, good for both queue and caching

**Alternatives Considered**:

| Option | Pros | Cons | Why Not Chosen |
|--------|------|------|----------------|
| RQ (Redis Queue) | Simpler API than Celery | Fewer features, less ecosystem | Need advanced features (priorities, chaining) |
| Dramatiq | Modern, better API than Celery | Smaller community, less mature | Celery's ecosystem worth the complexity |
| BullMQ (Node) | Excellent, modern | Requires Node backend | Using Python for ML libraries |
| Cloud tasks (GCP/AWS) | Managed service, no infrastructure | Vendor lock-in, cold starts | Local dev first |

**Decision**: Celery's maturity and feature set justify the learning curve.

---

## ML/Audio Technologies

### Source Separation: Demucs

**Chosen**: Demucs v4 (Meta Research)

**Why**:
- State-of-the-art audio separation quality (MDX leaderboard winner)
- 4-stem model (drums, bass, vocals, other) is good default
- 6-stem model available (drums, bass, vocals, guitar, piano, other)
- Open-source, MIT license
- PyTorch model, runs on GPU

**Alternatives Considered**:

| Option | Pros | Cons | Why Not Chosen |
|--------|------|------|----------------|
| Spleeter | Faster, lighter | Lower quality, no longer actively developed | Quality matters more than speed |
| X-UMX | Open-source, good quality | Slower than Demucs | Demucs quality worth extra time |
| commercial APIs | No GPU needed, better quality | Costly ($0.10+/song), privacy concerns | Local processing preferred for MVP |

**Decision**: Demucs offers best quality for a self-hosted solution.

---

### Transcription: YourMT3+ (Primary) + basic-pitch (Fallback)

**Chosen**: YourMT3+ (KAIST) with automatic fallback to basic-pitch (Spotify)

**Why YourMT3+**:
- **80-85% accuracy** vs 70% for basic-pitch
- State-of-the-art multi-instrument transcription model
- Mixture of Experts architecture for better quality
- Perceiver-TF encoder with RoPE position encoding
- Trained on diverse datasets (30k+ songs, 13 instrument classes)
- Open-source, actively maintained
- Optimized for Apple Silicon (MPS) with float16 precision (14x speedup)

**Why basic-pitch as Fallback**:
- Polyphonic transcription (multiple notes at once)
- Lighter weight, faster inference
- Simple setup, no model download required
- Good baseline quality (70% accuracy)
- Automatically used if YourMT3+ unavailable

**Alternatives Considered**:

| Option | Pros | Cons | Why Not Chosen |
|--------|------|------|----------------|
| MT3 (Music Transformer) | Google's latest, multi-instrument aware | Slower, larger model, harder to run | YourMT3+ more accurate |
| Omnizart | Multi-instrument, good documentation | Lower accuracy than YourMT3+, slower | Removed in favor of YourMT3+ |
| Tony (pYIN) | Excellent for monophonic | Only monophonic | Need polyphonic support |
| commercial APIs | Better quality | Expensive, privacy concerns | Local processing preferred |

**Decision**: YourMT3+ offers the best accuracy for self-hosted solution with intelligent fallback to basic-pitch for reliability.

---

## File Formats

### Primary Format: MusicXML

**Chosen**: MusicXML 4.0

**Why**:
- Industry-standard interchange format
- Supported by all major notation software (Finale, Sibelius, MuseScore, Dorico)
- Preserves notation semantics (clefs, articulations, lyrics)
- Human-readable XML (good for debugging)
- VexFlow can parse it directly

**Alternatives Considered**:

| Option | Pros | Cons | Why Not Chosen |
|--------|------|------|----------------|
| MIDI | Universal, compact, great for playback | No notation info (clefs, staff layout) | Complementary, not replacement |
| MEI (Music Encoding Initiative) | More expressive than MusicXML | Less tool support, steeper learning curve | MusicXML more widely adopted |
| ABC Notation | Human-readable text | Limited notation features, less standard | Better for folk music than general use |
| Proprietary (Finale .musx) | Native to notation software | Requires specific tools to read | MusicXML is open standard |

**Decision**: MusicXML is the universal standard for notation exchange.

---

### Intermediate Format: MIDI

**Chosen**: MIDI 1.0 (SMF Type 1)

**Why**:
- Universal output format from transcription models
- Easy to convert to MusicXML
- Useful for export option
- Tone.js plays MIDI directly

**Why Not Sufficient Alone**:
- Lacks notation semantics (clefs, key signatures, measure boundaries)
- No staff layout information
- Ambiguous rhythmic notation

---

## Development Tools

### Python Package Manager: uv or Poetry

**Chosen**: uv (recommended) or Poetry

**Why**:
- Reproducible builds with lock files
- Virtual environment management
- Faster than pip for large dependencies (PyTorch, etc.)

---

### Frontend Build Tool: Vite

**Chosen**: Vite

**Why**:
- Fast dev server with HMR
- Modern, best-in-class DX
- Great for React apps
- Smaller bundles than Webpack

---

### Containerization: Docker

**Chosen**: Docker + Docker Compose

**Why**:
- Consistent dev environment across machines
- Easy GPU passthrough for Demucs
- Simplifies Redis, API, worker orchestration

---

## Infrastructure (Future)

### Frontend Hosting: Vercel

**Recommended**: Vercel

**Why**:
- Excellent React/Vite support
- Global CDN
- Preview deployments for PRs
- Free tier is generous

**Alternative**: Netlify, Cloudflare Pages, AWS S3 + CloudFront

---

### Backend Hosting: Cloud Run or Modal

**Recommended**: Modal (for GPU workers)

**Why**:
- Serverless GPU containers
- Pay-per-use (no idle GPU cost)
- Fast cold starts
- Good Python support

**Alternative**: AWS ECS with GPU instances, GCP Cloud Run (CPU only, need separate GPU service)

---

### Database: PostgreSQL (future)

**Not needed for MVP** (using Redis for job state)

**When to add**:
- User accounts and auth
- Persistent job history
- Sharing features

---

## Decision Criteria Summary

When evaluating technologies, we prioritized:

1. **Quality Over Speed**: Better transcription/rendering > faster processing
2. **Open Source First**: Avoid vendor lock-in, control costs
3. **Python for ML**: Ecosystem too strong to ignore
4. **Standard Formats**: MusicXML/MIDI over proprietary
5. **Proven Tech**: Prefer mature libraries over bleeding edge
6. **Developer Experience**: Good docs and tooling matter

## Trade-off Examples

### Demucs vs. Spleeter
- **Chose Demucs**: Better quality worth 2x processing time
- **Rationale**: Users wait minutes anyway, quality is paramount

### VexFlow vs. OSMD
- **Chose VexFlow**: Editing capability > slightly better rendering
- **Rationale**: Users will edit output, need programmatic access

### FastAPI vs. Django
- **Chose FastAPI**: Async WebSocket support > admin panel
- **Rationale**: Real-time updates critical, don't need admin UI

## Next Steps

See [Deployment Strategy](deployment.md) for how these technologies deploy.