Justagwas
Prepare the recording
Decode the source, resample it to 44,100 Hz, and divide it into overlapping sections so boundary notes retain surrounding context.
A2M (Audio To MIDI)
Accurate local piano transcription with configurable CPU and GPU acceleration
Overview
A2M is a Windows desktop application designed to achieve high transcription accuracy when converting piano recordings into editable MIDI files. The A2M Piano Engine estimates notes, timing, velocity, sustain, and soft pedal events locally. Processing time depends on recording duration and available hardware, but transcription typically completes promptly on suitable systems.
- Best results from clean solo piano recordings
- Local analysis with no source recording upload
- Note onset, offset, velocity, sustain, and soft pedal estimation
- CPU processing by default, with optional CUDA or DirectML acceleration
- Separate CPU usage and GPU memory controls for balancing system load and speed
- Editable 960 PPQ MIDI output saved to Downloads/A2M by default
Download
Version 3.0.0 - July 14, 2026
Preview
Real UI captures from A2M.
How A2M turns sound into MIDI
A2M analyzes the acoustic signal itself. It prepares overlapping views of the recording, estimates coherent note and pedal intervals, refines their timing and velocity, and reconstructs the result on one MIDI timeline.
Construct spectral features
Transform centered audio frames into six complementary spectral views across 229 mel bands for model inference.
Decode musical events
Score possible note and pedal intervals, select a coherent path, and refine velocity and subframe timing.
Construct the MIDI file
Join section results on the original timeline, resolve overlaps, and write notes and optional pedal events at 960 PPQ.
Worked piano example
This 30 second solo piano excerpt pairs the source audio with a mapped score reference and the MIDI produced by A2M 3.0.0. The shared timeline makes pitch placement and onset timing directly inspectable.
Source recording
MP3, 44.1 kHz stereo, 30 seconds
282 of 282same pitch onsets matched within 250 ms
233 of 282same pitch onsets matched within 100 ms
66 msmean absolute onset difference
99.6%note accuracy in this example
Latest release and safety
Version 3.0.0 - July 14, 2026
- New transcription engine.
- Improved runtime reliability.
- Improved transcription accuracy.
- Updated CPU and GPU performance controls.
- Various other fixes.
Security and installation
Obtain A2M only from the official download links on this page. Windows SmartScreen may display a "Protected your PC" warning for a new or unsigned release.
The latest installer was scanned on VirusTotal with: 0/70 detections. View VirusTotal scan.
Manual installer and release details are available on the download page.
Open source and distribution
A2M is open source and distributed through GitHub and SourceForge. Source recordings remain on the local computer. Network access is used for requested model, runtime, or application downloads and for update checks when enabled.
