Measured 17 October 2026 · our own test

How accurate is TONIQ's audio to MIDI?

95.2 %

note F1 on solo piano, public test recordings

93.9 %

note F1 on acoustic guitar, public test recordings

A note counts as right when TONIQ writes the right pitch, within 50 cents, starting within 50 milliseconds of where it was played. F1 balances notes missed against notes invented. We ran this ourselves, on the build you download; the method and its limits are below, so you can judge it and repeat it.

The results.

Mean over files. Precision: of the notes TONIQ wrote, how many were real. Recall: of the real notes, how many TONIQ found. The last column also requires each note to end in the right place.

Test set What was tested Minutes Precision Recall Note F1 With note ends
Solo piano A public test set of recorded performances: 26 recordings 152.9 96.9 % 93.7 % 95.2 % 75.1 %
Acoustic guitar A public test set: 60 takes by one player, chords and solos 30.5 93.4 % 94.5 % 93.9 % 79.1 %

On guitar, chord accompaniment scored 93.3 % and solos 94.4 %. On piano the note-end score is measured against the notes as held by the sustain pedal.

How we measured it.

  • The build. The note detector that ships in the TONIQ installer, run with the Polyphonic engine at its default settings. Nothing tuned for the test.
  • The recordings. Two public test sets in which every note is known: piano performances captured note by note as they were played, and acoustic guitar takes with every note annotated. None of these recordings was used to build TONIQ.
  • The scoring. The standard open-source scoring for music transcription: onset within 50 ms, pitch within 50 cents, and for the note-end score, an end within 20 % of the note's length or 50 ms, whichever is larger.
  • Repeat it. Both test sets are public. Write to info@insektiq.com and we will send the scripts.

What it does not show.

  • Clean solo recordings. One instrument, well recorded. Dense mixes, heavy distortion and reverb are harder, for every tool.
  • Piano and guitar only. Vocals, drums and full mixes run on other engines in TONIQ and are not in these numbers yet.
  • Our own test. Not an independent review. We publish the method so that it can be checked.

Hear it on your own audio instead: TONIQ is free, with no length limit.

Questions about accuracy.

How accurate is TONIQ at audio to MIDI? +
On public test recordings that TONIQ never trained on, it found the right note at the right time with an F1 score of 95.2 % on solo piano and 93.9 % on acoustic guitar. We measured this ourselves on the shipping build on 17 October 2026; the method is on this page.
What does F1 mean here? +
Precision is the share of the notes TONIQ wrote that are really in the recording. Recall is the share of the real notes that TONIQ found. F1 combines the two into one number, so a tool cannot score well by writing too many notes or too few. A note counts as correct when its pitch is within 50 cents and its start is within 50 milliseconds of the real one.
Why is the note-end score lower? +
The second score also requires each note to end in the right place. Where a note ends is harder to define than where it starts: on piano the sustain pedal holds notes long after the key is released, and on guitar strings ring into each other. These are standard metrics, and transcription systems generally score lower on them.
Does this cover vocals, drums and full mixes? +
No. These two tests cover polyphonic piano and acoustic guitar with the Polyphonic engine. Vocals, drums and full mixes use other engines in TONIQ and are not part of these numbers. We will add them here when we have measured them the same way.
Is this an independent test? +
No. INSEKTIQ ran it. The test sets are public and the scoring is the standard one, so anyone can repeat it, and we will send the scripts on request.