Gamma Live

Try It

Quick Start

curl -X POST http://gamma.omegaai.dev/api/generate \
  -H "Content-Type: application/json" \
  -d '{"chords": "Dm7 G7 Cmaj7 Am7", "description": "smoky jazz bar, walking bass, sax solo"}' \
  --output song.mp3

curl -X POST http://gamma.omegaai.dev/api/generate \
  -H "Content-Type: application/json" \
  -d '{"chords": "C G Am F"}' \
  --output song.mp3

Returns an MP3 file. Generation typically takes 3–15 seconds depending on song length.

API Endpoints

Chord Format

Standard chord symbols are supported. Separate chords with spaces. Use newlines (or \n in JSON) to denote sections.

{"chords": "Am F C G\nAm F C G\nDm Am F E7"}

Styles

13 built-in styles, each with distinct instruments, drum patterns, and character. If you provide a description, GPT picks the best style automatically. Otherwise, the system auto-detects from chord content.

Error Handling

If the OpenAI call fails (rate limit, network error), the API does not return an error. It falls back to auto-detecting the style from chord content and applying a random variation. You always get music back.

How It Works

1. Description Interpretation (GPT-4.1-mini)

When you provide a description, it's sent to OpenAI's GPT-4.1-mini with a carefully engineered system prompt. The prompt encodes:

• All 13 styles with their character descriptions and BPM ranges
• All 26 variation parameters with exact numeric ranges and musical meaning
• Musical interpretation rules mapping vibes to parameters (e.g. "chill" → low drums, high humanize, laid-back groove)

GPT returns a JSON object selecting a style, BPM, and whichever variation parameters it deems relevant. The API then validates and clamps every value to its legal range before use — GPT never gets unchecked access to the generator.

If no description is provided, the system uses auto-detection: it analyzes the harmonic content of your chords (ratio of minor chords, jazz extensions, sus chords, power chords, etc.) and picks the most appropriate style with weighted randomization. A random variation with musical coherence (one of 8 "personality" archetypes like tight, loose, intimate, energetic, swung, cinematic, minimal, lush) is applied so each generation sounds unique.

2. Chord Parsing

The parser converts chord symbols into MIDI note numbers. It handles:

• Root detection — any note including sharps/flats (C, F#, Bb, etc.)
• Quality parsing — major, minor, diminished, augmented, sus, and all extensions (7, 9, 11, 13) with arbitrary alterations (#9, b13, etc.)
• Slash chords — Am/G puts G in the bass register while Am voices sit in the mid register
• Section breaks — newlines in the chord string create section boundaries, which affect dynamics, fills, and style morphing

Each chord becomes a bass note (placed in the 28–47 MIDI range) and a set of voicing notes (48–84 range), which are then distributed across tracks according to the style.

3. Multi-Track MIDI Generation

The core of the system. For each chord in the progression, up to 7 simultaneous tracks are generated:

4. The 26 Variation Parameters

These parameters control the musical character at a fine-grained level. GPT sets them from your description, or they're randomized coherently.

5. Audio Rendering

The generated multi-track MIDI file is rendered to audio in two stages:

• FluidSynth — Renders MIDI to WAV at 44.1kHz stereo using the FluidR3 GM soundfont (140MB, 128 General MIDI instruments). Each GM program number maps to a realistic instrument sound.
• FFmpeg — Converts WAV to MP3 using LAME at VBR quality 2 (~190kbps). The MP3 is streamed back as the HTTP response body, then temp files are cleaned up.

On Railway, rendering uses FluidSynth exclusively. Locally, the system can also use sfizz with higher-quality SFZ instrument libraries for per-track rendering with individual reverb and EQ, but these libraries are too large (4.3GB) for cloud deployment.

6. Architecture

• Framework — FastAPI (async, auto OpenAPI docs)
• Concurrency — Audio generation is CPU-bound (subprocess calls to FluidSynth/FFmpeg). Runs in a thread pool via asyncio.run_in_executor so the event loop stays responsive. A semaphore limits to 2 concurrent generations to prevent memory exhaustion.
• Temp files — Each generation creates MIDI, WAV, and MP3 in /tmp. After the MP3 streams back, a BackgroundTask cleans up all intermediate files.
• Timeout — 120 second hard limit per generation. Most requests complete in 3–15 seconds.
• Deployment — Docker container on Railway with Dockerfile builder. Health check on /health.
• Graceful degradation — If OpenAI is unavailable, the API still works. It falls back to chord-content-based style detection and coherent random variation.

7. Description Tips

The AI understands musical vocabulary and vibes. Here are some effective descriptions: