How accurate is audio-to-MIDI transcription? Real numbers, by instrument
Published benchmark data for the Basic Pitch model behind MelodyTrace — accuracy by instrument (vocals, guitar, piano, synths, orchestral), what the metrics mean, and how to get better results.
MelodyTrace's transcription runs on Basic Pitch, Spotify's open-source, instrument-agnostic audio-to-MIDI model. Instead of a "works great!" claim, here is what the model's own peer-reviewed benchmark reports about where it's strong and where it isn't — so you know what to expect before you upload, and which converter page fits your source material.
Where these numbers come from
These are not internal numbers — they are published results from the model's ICASSP 2022 paper, A Lightweight Instrument-Agnostic Model for Polyphonic Note Transcription and Multipitch Estimation (Bittner, Bosch, Rubinstein, Meseguer-Brocal, Ewert — Spotify / IRCAM), measured on standard public datasets with the MIREX note-transcription metrics from the mir_eval library:
- Acc — frame-level pitch accuracy: is the right note active at each 10 ms time step
- F (no offset) — the paper's main note-estimation metric: a note counts as correct if its pitch is within a quarter-tone and its onset lands within 50 ms of the reference, ignoring when the note ends
- F — the strictest metric: the same as above, but the note's offset (end time) must also land within 20% of its true duration
All three run 0–100%, higher is better.
Accuracy vs. instrument-specific tools
The paper also compares the general-purpose model against tools built for a single instrument — the more useful comparison if you're deciding whether a general converter is good enough for your source, or whether you need something specialized:
| Source | General model (Basic Pitch) | Instrument-specific tool |
|---|---|---|
| Solo guitar | Acc 71.7% · F-no-offset 84.0% · F 65.0% | TENT (guitar-only): Acc 63.2% · F-no-offset 76.3% · F 54.6% |
| Vocals | Acc 62.6% · F-no-offset 52.3% · F 34.6% | Vocano (vocal-only): Acc 61.6% · F-no-offset 64.2% · F 51.3% |
| Solo piano | Acc 37.5% · F-no-offset 70.9% · F 10.5% | Onsets & Frames (piano-only): Acc 43.8% · F-no-offset 95.2% · F 36.4% |
Three honest takeaways:
- Guitar is a genuine strength. The general-purpose model beats a dedicated guitar-transcription tool on every metric — see Guitar to MIDI.
- Vocals are close on raw pitch accuracy (62.6% vs. 61.6%) but behind on note segmentation (F-no-offset 52.3% vs. 64.2%). Expect the pitch to mostly land right, with some note boundaries needing manual correction around vibrato and slides — see Voice to MIDI and Humming to MIDI.
- Piano is the model's clearest weak point. A piano-specialized tool scores nearly 3.5x higher on the strict F metric (36.4% vs. 10.5%). If you're transcribing solo piano, budget real editing time in the piano-roll — see MP3 to Piano MIDI.
Accuracy across broader instrument categories
On the paper's full (not solo-only) test sets, the same model also reports:
| Instrument category | Acc | F (no offset) | F |
|---|---|---|---|
| Guitar (solo + ensemble) | 70% | 79% | 30% |
| Vocals | 63% | 52% | 35% |
| Orchestral | 53% | 49% | 35% |
| Synthesizers | 44% | 42% | 21% |
| Piano | 38% | 71% | 11% |
What this means for a file you upload
- Clean, monophonic sources — a single melody line, solo voice, solo guitar, lead instrument — transcribe closest to the numbers above. The model was evaluated on isolated stems, not full mixes.
- Dense polyphonic material (full band mixes, orchestral tutti) will produce more extra or missing notes than any of these benchmarks show, since real mixes combine instruments the model wasn't scored against in combination. See Song to MIDI for tips on isolating a part first.
- The piano-roll editor exists because of this data, not despite it. Even in the paper's best-case numbers, no instrument clears 100%. Moving, deleting, and re-transcribing with adjusted sensitivity is where you close the gap between the model's first pass and a usable score — not a workaround for a broken tool, but the expected second step for any AMT (automatic music transcription) system today.
Source
Bittner, R. M., Bosch, J. J., Rubinstein, D., Meseguer-Brocal, G., & Ewert, S. (2022). A Lightweight Instrument-Agnostic Model for Polyphonic Note Transcription and Multipitch Estimation. Proceedings of ICASSP 2022. arXiv:2203.09893 · github.com/spotify/basic-pitch