On the recordRead 12 Oct 202621 facts
BACH
On the record
- Architecture
- The architecture uses a BACH-1B decoder-only Transformer to produce ABC scores, with Dual-NTP and Chain-of-Score components.github.com · read 5 Oct 2026
- Control
- Its Dual-NTP component jointly predicts vocal and accompaniment patches at each step.github.com · read 12 Oct 2026
- Evaluation dependency
- The README says pitch extraction requires librosa, RMVPE, and its model checkpoint.github.com · read 5 Oct 2026
- Evaluation requirements
- Pitch extraction requires librosa, RMVPE, and its model checkpoint.github.com · read 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 · read 12 Oct 2026
- Intended users
- The README presents BACH for human-controllable long-song generation using bar-level symbolic notation.github.com · read 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 · read 12 Oct 2026
- Model
- The project describes BACH-1B as a decoder-only Transformer with 1 billion parameters.github.com · read 12 Oct 2026
- Model size
- The README describes BACH as a 1B-parameter model.github.com · read 5 Oct 2026
- Music representation
- The model uses non-overlapping 16-character patches per bar and outputs an ABC score.github.com · read 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 · read 12 Oct 2026
- Paper
- The repository says the related paper was accepted to ICASSP 2026.github.com · read 12 Oct 2026
- Preview contents
- The preview repository contains inference code and an example song, along with architecture figures.github.com · read 5 Oct 2026
- Product
- BACH is described as a human-editable, bar-level symbolic song generator.github.com · read 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 · read 5 Oct 2026
- Release status
- The repository calls itself a preview release containing partial inference, preprocessing, and evaluation utilities.github.com · read 12 Oct 2026
- Rendering tools
- The architecture description names MIDI, FluidSynth, and VOCALOID in the rendering and mixing flow.github.com · read 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 · read 5 Oct 2026
