Skip to content
Other

Dettato evidence repairer

error correction for music-to-MIDI transcription

2 min readupdated

This model serves as the second stage of KobiMusic's Dettato multi-instrument audio-to-MIDI (digital sheet music format) transcription system. It takes notes detected by the base dettato transcriber along with its internal states (feature vectors) and fixes transcription errors—removing false positives, shifting note onsets and offsets, and restoring missed notes. Built on a compact 4.9M-parameter transformer, it processes audio in 8-second windows, analyzing each note as an individual token. It is well suited for musicians, arrangers, and developers building tools for automated music transcription.

Key features

  • Internal feature awareness: alongside pitch, velocity, and duration, the model reads a 769-value vector containing onset/offset probabilities and hidden states from the base transcriber.
  • Fine-grained timing adjustments: predicts note retention probability, shifts onsets in 5 ms steps within a ±0.5 s window, and refines note durations.
  • Missed note recovery: evaluates two insertion slots per note, adding missed pitches at a probability of 0.6 or higher while filtering out duplicates within 30 ms.
  • Lightweight architecture: 9 layers with a hidden dimension of 192 and 6 attention heads, using rotary position embeddings over onset time and pitch.

What you can do

  • Clean up raw transcriptions: pair it with the base dettato checkpoint to strip out false positives and sharpen note endings in raw MIDI output.
  • Run the full Dettato pipeline: invoke it via the dettato CLI to automatically split audio into MIDI tracks, followed by instrument grouping (dettato-grouper) and drum separation (dettato-drums).

Example

from dettato.audio import load
from dettato.evidence import EvidenceRepairer
from dettato.transcriber import Transcriber

t = Transcriber.from_pretrained("cuda")
rep = EvidenceRepairer.from_pretrained("cuda")
notes, original, info = rep(t, load("song.mp3").at(16000))
Read moreDetails
Benchmarks (9 datasets, including Slakh2100): overall composite score improved from 56.65 to 56.95 (+0.30; onset gain +0.38, offset gain +1.96, frame score -1.14 due to shortened sustained notes).
Limitations: handles pitched notes only without instrument assignments, requires the base dettato checkpoint, and passes windows denser than 4,096 notes per 8 s unchanged.
Training and license: trained on internal synthetic data with no benchmark leakage; released under the Apache-2.0 license.
Examples
Discuss Telegram 0 570

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 0.0 GB · ready files0.5 GB
  • BF16weights 0.0 GB0.5 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · F32 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # kobimusic/dettato-evidence
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download kobimusic/dettato-evidence --local-dir ~/models/dettato-evidence
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to dettato-evidence
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of dettato-evidence (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

kobimusic/dettato-evidence
Author
kobimusic
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

5700

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Other

DAVID-Net

DAVID-Net