Spaces:
Running on Zero
A newer version of the Gradio SDK is available: 6.22.0
title: MERIT
emoji: 🎚️
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.28.0
python_version: '3.12'
app_file: app.py
pinned: false
license: cc-by-nc-4.0
short_description: Compare melody, rhythm, and timbre similarity
MERIT Music Similarity
A HARP-compatible deployment of MERIT for comparing two music clips along three independently learned dimensions: melody, rhythm, and timbre.
The Space uses the official MERIT projection heads with the frozen MERT-v1-330M backbone. Model files are downloaded from their official Hugging Face repositories at pinned revisions and cached between requests.
Inputs
- Two music audio clips
- Between 5 and 10 seconds per clip
Audio is converted to mono, resampled to 24 kHz, and padded or truncated to 10 seconds. The MERIT projection heads were trained from 10-second segments; very short inputs are rejected because silence padding can make their scores misleadingly high.
Output
A JSON file containing cosine similarity scores for:
- Melody
- Rhythm
- Timbre
Scores range from -1 to 1. They are similarity measurements, not calibrated probabilities.
Sources
See SOURCES.md for upstream code, model revisions, licenses, and the associated research papers.
License
The MERIT code and projection heads are released under the MIT License. The MERT-v1-330M backbone is released under CC BY-NC 4.0. This combined deployment is therefore intended for non-commercial use.