Downloads · 30 days
350
100% of all-time downloads
mt3-GGUF
cstr/mt3-GGUF
mt3-GGUF is a audio classification model from cstr. Use it for the audio classification task on the model card, and read the license before you ship it in a product. It is set up for ggml. The card lists the license as apache-2.0.
GGUF conversion of Google Magenta's MT3 multi-instrument music transcription model (Apache-2.0) for CrispStrobe/CrispASR (§250).
All-time downloads
350
Public
Repo size
96 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.gguf96 MB · 100%
At a glance
- Task
- Audio Classification
- Library
- ggml
- License
- apache-2.0
- Access
- Public
- Created
- Sep 2, 2026
- Updated
- Sep 2, 2026
- SHA
- 4299d5f3
Base models
Model card
From the Hugging Face model README
MT3 — GGUF (ggml)
GGUF conversion of Google Magenta's MT3 multi-instrument music transcription model (Apache-2.0) for CrispStrobe/CrispASR (§250).
MT3 is a T5-style encoder/decoder that reads a log-mel spectrogram and emits MIDI-like event tokens, transcribing multiple instruments at once — every note carries a General MIDI program, and drums are a separate class. It is the only per-program note-event model in CrispASR (piano-transcription is 88-key piano, basic-pitch is single-track polyphonic).
Files
| File | Size | Notes |
|---|---|---|
mt3-f16.gguf | 96 MB | 190 tensors; encoder+decoder, sinusoidal absolute positions, event codec |
Converted with models/convert-mt3-to-gguf.py, which decodes the T5X/zarr checkpoint with stdlib + numpy only (no JAX, t5x or TensorStore). Weights verified bit-exact against the reference PyTorch export (189/189 tensors, max abs diff 0.0).
Usage
crispasr --piano --backend mt3 -m auto --auto-download -f song.wav
crispasr --piano -m mt3-f16.gguf -f song.wav --piano-format json # + program/instrument/is_drum
crispasr --piano -m mt3-f16.gguf -f song.wav --piano-format midi # multi-track SMF
--piano-format midi writes a format-1 Standard MIDI File with one track per program and drums on GM channel 10.
Parity
Validated stage-by-stage against a numpy reference implementation of the upstream graph (tools/reference_backends/mt3.py):
| stage | result |
|---|---|
| mel front end | cos 1.000000000 |
| encoder output | cos 0.999999879 |
| first-step decoder logits | cos 0.999999999, argmax identical |
| greedy tokens | identical, 22/22 steps |
| end-to-end note events | exact positional match (88/88, 5/5, 8/8 across three clips) |
The GGUF records mt3.pos_embed=sinusoidal and mt3.use_relative_attention_bias=0; the runtime hard-fails rather than silently falling back to T5's relative-bucket attention, which is a known failure mode of other ports.
- Task
- Audio Classification
- Library
- ggml
- License
- apache-2.0
- Created
- Sep 2, 2026
- Updated
- Sep 2, 2026