Guitar Audio to MIDI Test: Solos, Chords and Missing Notes

What did we test?

These are real acoustic-guitar performances from GuitarSet v1.1.0: bossa nova and jazz solos, plus funk and rock comping. The input is the mono microphone recording, not a six-channel hex pickup, a synthesized example or a complete commercial song. A mono recording can still contain several simultaneous pitches.

Two development excerpts and two holdout excerpts were fixed before viewing model results. The holdout tracks are retained even when a research configuration performs worse. Each complete file was used; the table identifies the real duration and reference-note count.

How is one correct note defined?

The main metric requires pitch within 50 cents and onset within 50 ms. Maximum one-to-one bipartite matching pairs each event at most once. Remaining reference events count as misses; remaining predictions count as extras. F1 is 2 × hits / (2 × hits + extras + misses). Totals are micro-averages, computed from all event counts rather than an average of track percentages.

Offset-aware F1 additionally requires the ending within max(50 ms, 20% of the reference duration). References keep continuous sounding pitches and per-string unisons; they are not rounded, deduplicated or octave-shifted to help the score. Predictions were not manually edited or shifted in time.

GuitarSet onsets were checked by its authors, while offsets and pitch were estimated semi-automatically. Reference labels can still contain errors. Per-string multiplicity can also exceed what a pitch-only model can identify. Confidence was not supplied in the extracted reference events; its original missing values remain in the downloads.

Results, including the holdout group

Across all 651 reference notes, the earlier standard produced 606 events and the standard candidate 608. Their F1 values were 71.76% and 71.80%. This small change is not evidence that transcription generally became more accurate. One development comping excerpt regressed; the per-excerpt table keeps that result.

The standard changes target end-window handling and sample-rate processing. These four 44.1 kHz recordings do not independently establish high-sample-rate accuracy. Measurements bind the exact research script hashes in the downloadable data, not every later build or device.

Why the short-note setting is not the default

The rapid research preset increased recall for notes shorter than 128 ms from 31.76% to 79.41% across these excerpts. It also increased extra notes from 156 to 246. More output notes do not mean more correct notes.

On the fixed holdout group, standard F1 was 76.77% and rapid F1 76.50%. The research preset did not beat standard on this group and was not enabled as the website default. A fast tempo alone is not a sufficient reason to loosen note filtering.

What should you correct by hand?

Check extra notes, missed chord tones, onsets and endings separately. An octave doubling can be real music or a false harmonic detection; never delete all upper-octave notes automatically. Different string attacks inside a strum can also be intentional, so snapping them all to one instant can erase the performance.

An automatic screen found 8 possible octave mismatches for standard and 10 for rapid. It only pairs otherwise-unmatched events near an onset and 12 or 24 semitones apart; it is not a human-confirmed error count.

Keep the raw result read-only, create a correction branch and compare the same passage with the original recording. Change confirmed pitches and times, then export the fixed MIDI version. Tempo, meter and a readable notation grid need a separate review.

Listen and reproduce the test

The complete Rock input is mirrored below. Audio does not preload: downloading starts only when you play it or choose its download link. Other recordings remain available from the original dataset. All four full references and all three prediction configurations are provided without manual note edits.

Download unchanged Rock WAV · 2,288,022 bytes

• Full metrics, event matches and script identities (JSON)
• Attribution and changes
• File checksums

In the workspace, select the same local WAV and its full time range, retain the raw result and compare event data. The benchmark used Windows Chrome 156.0.8078.79 with one Worker at a time. These are test-host observations, not performance promises for Macs, phones or every browser.

Limits of this result

Four guitar excerpts are a small diagnostic cohort. They do not establish accuracy for vocals, drums, bass mixes, piano, distortion or a full commercial song. No human listening or playability acceptance is claimed. Successful file generation and a high note score do not establish a readable score or correct guitar fingering.

The local model does not output pitch-bend curves. Instrument separation, guitar string/fret assignment and automatic chord verification are not demonstrated here. Your own recording still needs review.

Dataset credit and reuse

GuitarSet v1.1.0 by Qingyang Xi, Rachel M. Bittner, Johan Pauwels, Xuzhou Ye and Juan P. Bello (2019), DOI 10.5281/zenodo.3371780. Dataset audio and annotations are provided under CC BY 4.0. The mirrored Rock WAV is the complete, unchanged source recording: no cropping, normalization or transcoding. Reference JSON retains all extracted string-note events; predictions are generated analysis, not manual corrections. No endorsement is implied.

Original GuitarSet record · CC BY 4.0 · Dataset recording and annotation method