# Sources

## MERIT

- Paper: https://arxiv.org/abs/2605.27346
- Source: https://github.com/AMAAI-Lab/MERIT
- Projection heads: https://huggingface.co/amaai-lab/merit
- Pinned head revision:
  `a85df30eca1ba112eb594285f3ba1d96488e7883`
- Code and projection-head license: MIT

## MERT

- Paper: https://arxiv.org/abs/2306.00107
- Backbone: https://huggingface.co/m-a-p/MERT-v1-330M
- Pinned backbone revision:
  `5240c2708a5acaee1007f43fb9735c7dcd0b78c9`
- Model license: CC BY-NC 4.0

The deployed inference procedure follows the official MERIT training code:
audio is converted to mono at 24 kHz, padded or truncated to the 10-second
segment duration used to extract the training embeddings, encoded with MERT
layers 3, 4, 5, 6, and 23, mean-pooled and concatenated, then projected by
the three factor-specific heads.