MuScriptor β€” ONNX (small & medium, INT8)

ONNX Runtime conversions of MuScriptor Small (β‰ˆ100M params) and MuScriptor Medium (β‰ˆ300M params), the multi-instrument music transcription model by Mirelo Γ— Kyutai. Made for running the model on CPU through ONNX Runtime (e.g. onnxruntime-node) without PyTorch. Format conversion only β€” nothing retrained. See NOTICE.md for attribution, the list of changes and the upstream conditions of use.

License: CC BY-NC 4.0 β€” non-commercial use only, plus the upstream conditions of use in NOTICE.md. By using these files you accept them.

Files

File Size What
small/cond.onnx 23.6 MB conditioner: audio β†’ prefix embeddings (log-mel inside the graph)
small/lm-int8.onnx 108.5 MB decoder, INT8 weights, KV cache in/out
medium/cond.onnx 25.1 MB conditioner
medium/lm-int8.onnx 315.2 MB decoder, INT8 weights
SHA256SUMS checksums of the four graphs
tools/ export + quantization scripts that produced these files

Graph interface (both sizes)

cond.onnx

audio f32 [1, 80000] β€” one 5 s chunk at 16 kHz mono, zero-padded
instrument i64 [1, K] β€” MT3_FULL_PLUS instrument group ids; -1 = unconditional
dataset i64 [1, 1] β€” always -1
β†’ prefix f32 [1, 501 + 1 + K, dim]

The mel spectrogram (reflect pad, STFT as a DFT convolution, HTK mel bank, log(x + 1e-6)) runs inside the graph and uses the checkpoint's own window and filterbank buffers, which differ from the textbook formulas by about 2.4e-4.

lm-int8.onnx

tokens i64 [1, S]
prefix f32 [1, P, dim] β€” conditioner output on the first call; P = 0 afterwards
past_k, past_v f32 [L, 1, T_past, H, 64] β€” empty (T_past = 0) on the first call
β†’ logits f32 [1, card] β€” last position only
β†’ present_k, present_v the cache grown by this call
dim heads (H) layers (L) card (vocab) initial token
small 768 12 14 1393 1393
medium 1024 16 24 1395 1395

Positions continue from T_past; the same graph serves prefill and single-token decoding. Decoding is greedy over the first 1393 logits, 5 s chunks, MT3 token vocabulary, with the tie-prologue handling described in the reference implementation. Prefill the first call with [initial_token] (plus a forced tie prologue for chunks after the first).

Fidelity

fp32 exports vs the PyTorch reference, same 5 s chunk: conditioner output max |Ξ”| 5.7e-5 (small) / 7.1e-5 (medium); greedy token sequences identical (24/24 and 27/27 tokens).

INT8 decoders, teacher-forced top-1 agreement with the fp32 reference over the reference's greedy tokens (5 short musical clips, 175 tokens for small and 216 for medium β€” a small sample, not a benchmark):

fp32 INT8 (these files)
small 100% 96.6%
medium 100% 98.6%

INT4 weight-only variants were also tried (small 96.6%, medium 94.0%) and are not published here.

Reproducing

python tools/export_muscriptor_onnx.py --size small --checkpoint model.safetensors --reference <muscriptor repo> --out out/small
python tools/quantize_int8.py out/small/lm.onnx out/small/lm-int8.onnx --card 1393

The upstream weights are gated on Hugging Face: accept their conditions on the model pages first.

Citation

@misc{rouard2026muscriptoropenmodelmultiinstrument,
  title={MuScriptor: An Open Model for Multi-Instrument Music Transcription},
  author={Simon Rouard and Michael Krause and Axel Roebel and Carl-Johann Simon-Gabriel and Alexandre D\'efossez},
  year={2026},
  eprint={2607.08168},
  archivePrefix={arXiv},
  primaryClass={cs.SD},
  url={https://arxiv.org/abs/2607.08168}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for fpp-ableton/muscriptor-onnx

Quantized
(4)
this model

Paper for fpp-ableton/muscriptor-onnx