FREEDOM251

Freedom251 › AI Music Generators

BACH

Paid Self-hosted

Visit BACH Alternatives

BACH's own site

What BACH is

BACH is an open-source, self-hosted system for generating songs as editable, bar-level symbolic scores. Its described workflow uses an LLM to create lyrics, then a Transformer to produce an ABC score that rendering tools turn into music. The BACH-1B model is a decoder-only Transformer with 1 billion parameters. It uses Dual-NTP to predict vocal and accompaniment patches together and Chain-of-Score section tags, including chorus markers, to support longer song structures. Scores use non-overlapping 16-character patches per bar. The documented rendering pipeline converts ABC to MIDI and names FluidSynth and VOCALOID in the stereo-mix stage. The repository is a preview release with partial inference, preprocessing, and evaluation utilities, plus an example song and architecture figures. It is not a complete CPU-ready generation pipeline, and required codec modules and assets are absent. Pitch extraction also requires librosa, RMVPE, and its checkpoint. The README recommends trusted model assets and describes restricted weights-only checkpoint loading.

Who it suits

BACH suits people seeking human-controllable long-song generation through bar-level symbolic notation. Its self-hosted preview is relevant to users prepared to work with a partial release and its listed dependencies.

What it does well

  • Open-source and self-hosted.
  • Joint vocal and accompaniment patch prediction.
  • Section tags support long-form structure.
  • Includes inference code and an example song.

What to know first

  • Preview is not a complete CPU-ready pipeline.
  • Required codec modules and assets are absent.
  • Pitch extraction needs librosa, RMVPE, and a checkpoint.

Verdict

BACH describes a structured approach to editable song generation, with joint vocal and accompaniment output and section-level controls. The preview release has substantial pipeline gaps, so it is not presented as a ready-to-run CPU generation system.

Facts from BACH's own pages

Architecture
The architecture uses a BACH-1B decoder-only Transformer to produce ABC scores, with Dual-NTP and Chain-of-Score components. · github.com, 5 Oct 2026
Control
Its Dual-NTP component jointly predicts vocal and accompaniment patches at each step. · github.com, 12 Oct 2026
Evaluation dependency
The README says pitch extraction requires librosa, RMVPE, and its model checkpoint. · github.com, 5 Oct 2026
Evaluation requirements
Pitch extraction requires librosa, RMVPE, and its model checkpoint. · github.com, 12 Oct 2026
Generation flow
The README says an LLM writes lyrics, a Transformer emits an ABC score, and off-the-shelf renderers produce music. · github.com, 12 Oct 2026
Intended users
The README presents BACH for human-controllable long-song generation using bar-level symbolic notation. · github.com, 12 Oct 2026
Maker affiliation
The repository owner's profile identifies Tongxi Wang as affiliated with the School of Future Technology at Southeast University. · github.com, 12 Oct 2026
Model
The project describes BACH-1B as a decoder-only Transformer with 1 billion parameters. · github.com, 12 Oct 2026
Model size
The README describes BACH as a 1B-parameter model. · github.com, 5 Oct 2026
Music representation
The model uses non-overlapping 16-character patches per bar and outputs an ABC score. · github.com, 12 Oct 2026
Notable limit
The preview is not a complete CPU-ready generation pipeline, and required codec modules and assets are absent from the repository. · github.com, 12 Oct 2026
Paper
The repository says the related paper was accepted to ICASSP 2026. · github.com, 12 Oct 2026
Preview contents
The preview repository contains inference code and an example song, along with architecture figures. · github.com, 5 Oct 2026
Product
BACH is described as a human-editable, bar-level symbolic song generator. · github.com, 12 Oct 2026
Release limitation
The README says the preview is incomplete and not a CPU-ready generation pipeline; required codec modules and assets are not included. · github.com, 5 Oct 2026
Release status
The repository calls itself a preview release containing partial inference, preprocessing, and evaluation utilities. · github.com, 12 Oct 2026
Rendering tools
The architecture description names MIDI, FluidSynth, and VOCALOID in the rendering and mixing flow. · github.com, 12 Oct 2026
Score structure
The README says the model predicts vocal and accompaniment patches jointly and uses section tags for long-form coherence. · github.com, 5 Oct 2026
Security guidance
The README advises using trusted model assets and says inference uses PyTorch's restricted weights-only checkpoint loader rather than automatically loading legacy pickled model objects. · github.com, 12 Oct 2026
Song structure
Chain-of-Score uses section tags such as chorus start and end markers to support long-form coherence. · github.com, 12 Oct 2026
Vocal generation
Yes · github.com, 22 Sept 2026

Where BACH stands on Freedom251

Look at these too