Mirelo transforms a finished track into separate MIDI tracks

Audio-to-MIDI Pro transcribes complete songs of up to 15 minutes into MIDI tracks by instrument, sheet music, chords, and editable MusicXML files.

Sometimes all that remains of a song is a stereo file. The original tracks are gone, the project no longer opens, or the music comes from a recording for which no one has the stems. Audio-to-MIDI Pro attempts to work backward from that final mix and recover the notes played by each instrument.

Mirelo’s service accepts tracks up to 15 minutes long. It detects the instruments present, assigns notes to them, and creates a separate MIDI track for each one. It also produces a score, a chord chart, and information about the key, tempo, and time signature.

The result should not be confused with audio source separation. Audio-to-MIDI Pro does not recover the original guitar, drum, or vocal recording. It reconstructs a symbolic representation of the performance, including pitches, note start and end times, their placement in time, and the instrument assigned to them.

A MIDI track therefore does not contain the sound captured in the recording. An exported bass line must be played through a virtual instrument or compatible hardware. This separation between the notes and their sound is precisely what makes it useful for production: musicians can change the instrument, transpose a part, move a note, alter the rhythm, or reorchestrate the entire piece.

The difficulty is that the service works directly from the complete mix. Several instruments may produce similar frequencies at the same time, while reverb, distortion, compression, and other effects can blur their identities. A doubled guitar, a synthesizer sitting behind a vocal, or a bass following the kick drum can be difficult to distinguish without the original tracks.

Mirelo nevertheless says its system processes the entire arrangement in a single operation, while many transcription tools specialize in isolated piano, voice, guitar, or drums. Each detected part appears on its own track in Mirelo Studio and can be edited in a piano roll.

Users can compare the original recording with the MIDI reconstruction, isolate an instrument, assign it a reference sound, and listen to individual notes while editing them. A metronome follows the detected beats, making it easier to check whether the transcription remains aligned with the performance.

Audio-to-MIDI Pro is not limited to notes. It estimates chords, key, tempo, tempo changes, time signature, and the positions of beats and downbeats. This information can be preserved when the result is imported into a digital audio workstation.

The export system offers two approaches to timing. The first preserves variations in the performance through a tempo map. It is intended for songs that speed up, slow down, or were not recorded to a click. The second brings the transcription to a fixed tempo and can quantize the notes to the detected beats.

This correction makes editing easier, but it necessarily changes part of the performance. Intentional placement behind the beat, swing, or expressive timing variations may be treated as inaccuracies. The choice between preserving the original feel and aligning notes to a grid therefore depends on the intended use.

The score is generated from the same data. Mirelo selects clefs, adjusts octave placement, and uses percussion notation for drums. The service can export a full score or separate parts as PDF and MusicXML files. Tablature is also available for supported guitar and bass parts.

PDF files can be viewed and printed without notation software. MusicXML allows users to continue editing in a compatible application. Changes made in the piano roll are reflected in the score, avoiding the need to correct the two representations separately.

An automatically generated score does not become a definitive edition, however. Readability depends on the accuracy of the detected tempo, time signature, quantization, and note durations. Ties, polyphonic voices, triplets, time-signature changes, and certain articulations may require manual revision before rehearsal or publication.

Chord detection has also been revised. Mirelo says the new version prefers to leave some passages unlabeled rather than add too many uncertain chords. This selectivity may make the chart easier to read, but no separate evaluation currently measures the error rate, particularly for inversions, extended chords, passing notes, or deliberately ambiguous harmony.

Audio-to-MIDI Pro grew out of work conducted with Kyutai and researchers from IRCAM. An initial version called MuScriptor was introduced in July 2026, along with published code and weights. The commercial version offered in Mirelo Studio uses a model described as more accurate and trained on additional data.

This distinction matters. The publicly available weights belong to MuScriptor and do not necessarily match the production model used by Audio-to-MIDI Pro. Mirelo still links to the repository from the product page, but it does not provide a download specifically identified as the new Pro version. That version remains accessible through Studio and the API.

MuScriptor receives short audio segments and produces a sequence of elements representing time, pitch, and instrument type. The successive transcriptions are then assembled to reconstruct the song. An optional indication of the expected instruments can help stabilize their assignment across segments.

The research paper describes pretraining on approximately 1.45 million MIDI files converted into synthetic recordings. It then mentions 170,000 real-world recordings totaling more than 11,000 hours, accompanied by note annotations, followed by an additional stage using 300 songs with more carefully reviewed transcriptions.

Three sizes of the research model are available. They contain approximately 103 million, 307 million, and 1.4 billion parameters. The smallest can run on a CPU, the middle version is presented as the default compromise, and the largest is better suited to a graphics card.

The weights are distributed under a CC BY-NC 4.0 license, which permits their use and modification with attribution but prohibits