---
license: apache-2.0
language:
- en
tags:
- music
- midi
- audio-to-midi
- piano-transcription
- aria-amt
- automatic-music-transcription
---

# Aria-AMT Mirror

Mirror of the [EleutherAI Aria-AMT](https://github.com/EleutherAI/aria-amt)
piano transcription checkpoint (`piano-medium-double-1.0.safetensors`),
re-hosted from [loubb/aria-midi](https://huggingface.co/datasets/loubb/aria-midi)
for use in the [MAESTRO AI Workstation](https://github.com/AEmotionStudio).

## What this model does

**Audio → MIDI piano transcription.** Takes an audio recording of a piano
performance (mp3, wav, flac) and outputs a MIDI file with notes, timings,
velocities, and pedal events. State-of-the-art quality for solo piano —
comparable to or better than Magenta's classic Onsets-and-Frames.

## Architecture

Sequence-to-sequence transcription model. Variant: `medium-double` —
the standard Aria-AMT v1.0 release.

## License

**Apache-2.0** — commercial-use OK.

## Usage in MAESTRO

Loaded by `backend/ai/models/aria_amt.py`; surfaced in the AI Workstation's
`TranscribePanel` under the "Piano" mode tab.

## Citation

```
@misc{bradshaw2024ariaamt,
  title={Aria-AMT: Efficient and robust automatic piano transcription},
  author={Louis Bradshaw and EleutherAI},
  year={2024},
  url={https://github.com/EleutherAI/aria-amt},
}
```