Static AI Studio

Static AI Studio is a local, artist-authorized AI vocal system. Its working proof-of-concept pipeline generates a rap performance from typed lyrics with ACE-Step 1.5, converts that performance into Static's voice with RVC, and targets a clean, dry WAV that Static can edit and mix.

Static has authorized the use of his recorded vocals for this project. Only authorized recordings and models should be added to this repository.

Scope

This repository is intended to support:

• preparing authorized, isolated vocal recordings;
• training and evaluating a voice model for Static;
• preserving the timing, phrasing, cadence, and expression of a new performance;
• local inference that exports clean, dry WAV vocals.

The current checkpoint covers the local AI pipeline, not the planned Prompt Lab UI. Mastering and final-song mixing remain outside the project.

Repository layout

engines/
  voice/
    preprocess/   Vocal cleanup, segmentation, and dataset preparation
    training/     Voice-model training code
    inference/    Performance-to-voice conversion and dry WAV export
    models/       Local model artifacts (ignored by Git)
  performance/    Typed-lyrics performance generation and conditioning
data/
  static/
    raw/vocals/   Original authorized vocal recordings
    processed/    Cleaned and segmented material
    training/     Training-ready manifests and features
    references/   Approved evaluation/reference clips
projects/songs/   Per-song inputs, outputs, and settings
scripts/          Local utilities
configs/          Versioned configuration files
docs/             Architecture and delivery roadmap

Empty working directories contain .gitkeep placeholders. Audio, datasets, model weights, and generated song files are excluded from Git by default.

Current status

The disposable static-test-v0 proof of concept is trained and indexed with RVC v2 at 40 kHz. ACE-Step 1.5 Turbo can generate typed-lyrics rap performances on the RTX 2070 using FP32 diffusion, TorchAO INT8, aggressive CPU offload, and tiled CUDA VAE decoding. Environments, models, authorized audio, training artifacts, and generated WAV files remain local and ignored.

See docs/PIPELINE.md for pinned revisions, patch application, environment boundaries, and the proven inference configuration. The machine-readable experiment record is configs/static-test-v0.yaml.

Check the host system

The system check uses only Python's standard library and read-only operating-system commands:

python3 scripts/check_system.py

Missing tools are reported as unavailable; they are not installed automatically.

Recommended input and output contract

• Input: a clean, dry, mono rap performance, ideally PCM WAV with minimal room sound.
• Output: a clean, dry PCM WAV aligned to the source performance.
• Preserve: lyrics, timing, cadence, articulation, pitch movement, dynamics, and breaths where practical.
• Change: vocal identity/timbre toward Static's authorized target voice.

Technical choices, sample rates, model families, and training settings remain intentionally undecided until the machine and authorized dataset have been audited.