Neural sound-texture foundry for scorekit powered by Magenta RealTime 2 (MRT2).
scorekit-texture-forge compiles declarative sound-texture recipes into certified, ready-to-consume textures.yaml profiles + 48kHz stereo WAV assets conforming strictly to scorekit's M14 discovery contract.
recipe.yaml ──► texture-forge build ──► textures.yaml + *.wav ──► scorekit build --texture-profile ...
scorekit is a deterministic game music compiler with an iron rule: no in-house creative DSP or neural generation inside the compiler.
However, video games frequently require atmospheric texture beds (space station engine drones, subterranean room tones, alien winds, or mechanical impacts) that cannot be synthesized by General MIDI or standard acoustic SoundFonts.
scorekit-texture-forge acts as the external asset foundry (analogous to scoredata-forge for SFZ instruments):
- Generative Neural Audio: Uses Magenta RealTime 2 (MRT2) to synthesize organic, complex, and immersive textures from text prompts.
- Strict Discovery Contract: Automatically structures and certifies the output into
textures.yamlwith explicit categories, tags, playback constraints, and SHA-256 checksums. - Deterministic Assembly: Passes the resulting library to
scorekit, which handles sample-exact timeline placement, seamless loop sealing, and stem alignment.
- Declarative Recipes: Define your sound library in concise YAML (
recipes/*.yaml). - Magenta RealTime 2 Integration: Native support for Apple Silicon MLX, JAX, and C++ CLI (
hello_mrt2). - Deterministic Synthetic Fallback: Built-in
mockengine for instant local prototyping and CI testing without multi-gigabyte neural checkpoints. - scorekit M14 Compliance: Emits
schema_version: 1profiles, verified directly viascorekit texture check. - Agent-Ready: Machine-readable
--jsonmodes, clean exit codes, and a bundled Agent Skill for autonomous composing workflows.
# Clone the repository
git clone https://github.com/talkincode/scorekit-texture-forge.git
cd scorekit-texture-forge
# Standard installation (supports CLI & synthetic mock engine)
pip install -e .
# With Magenta RealTime 2 support (Apple Silicon / MLX)
pip install -e ".[mrt2]"# Build from a bundled recipe
texture-forge build recipes/scifi-station.yaml -o out/scifi --verify --jsonOutput directory structure:
out/scifi/
├── textures.yaml # scorekit M14 texture profile
├── reactor_hum.wav # 48kHz 16-bit PCM stereo audio
├── airlock_hiss.wav # 48kHz 16-bit PCM stereo audio
├── SHA256SUMS # Cryptographic checksums
├── generator.json # Provenance and physical metrics
└── README.md # Library usage documentation
texture-forge inspect recipes/scifi-station.yamltexture-forge lint out/scifi --jsonIn your scene.yaml:
title: Abandoned Station Core
tempo: 90
key: C_minor
time_signature: "4/4"
bars: 8
loop: true
textures:
# Continuous ambient loop
- source: reactor_hum
mode: loop
gain: 0.35
# Triggered event scheduled on beat 4 and beat 20
- source: airlock_hiss
mode: one_shot
at: [4.0, 20.0]
gain: 0.5
tracks:
- id: pad
instrument: choir_pad
pattern: sustain
intensity: 0.5
- id: bass
instrument: synth_bass
pattern: bass
intensity: 0.6Build game-ready audio assets:
scorekit build scene.yaml --texture-profile out/scifi/textures.yaml -o station.ogg --stemsThis repository includes an Agent Skill at skills/scorekit-texture-forge/ that teaches AI coding assistants how to:
- Formulate sound recipes adhering to the 8 closed categories (
ambience,foley,impact,transition,tonal,industrial,organic,sound_design). - Trigger automated forge builds and inspect validation errors.
- Seamlessly wire generated texture profiles into game scoring scenes.
Install the skill for Claude Code / Antigravity / Cursor:
cp -r skills/scorekit-texture-forge ~/.claude/skills/MIT License. See LICENSE for details.