CAL-MOS · W2V BERT 2.0 · TMHINT-QI

CAL-MOS is a non-intrusive Mean Opinion Score (MOS) predictor built on speech foundation models. This checkpoint uses the Adapters + Mean (A+M) configuration with W2V BERT 2.0, trained on TMHINT-QI.

Paper · arXiv · GitHub

Model

CAL-MOS collects representations across the frozen encoder depth, calibrates each layer with a lightweight adapter, and combines the adapted representations before MOS regression.

Usage

Load the model directly from the Hugging Face Hub with Transformers. No clone or manual snapshot download is required.

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "alefiury/CALMOS-W2V-BERT-2.0-TMHINT-QI",
    trust_remote_code=True,
).to("cuda")

mos = model.predict("audio.wav")
print(mos)

Batch inference is also supported:

scores = model.predict(
    [
        "audio_1.wav",
        "audio_2.wav",
        "audio_3.wav",
    ],
    batch_size=8,
)

print(scores)

Audio is converted to mono and resampled to 16,000 Hz automatically. Predictions are clipped to the [1, 5] MOS range by default.

trust_remote_code=True is required because the lightweight CAL-MOS architecture is shipped with this model repository. The frozen backbone is downloaded automatically from its original Hugging Face repository.

Citation

@inproceedings{ferreira26_interspeech,
  title     = {{CAL-MOS: Bridging Layers with Adapters for Robust MOS Prediction Across Speech Foundation Models}},
  author    = {Alef Iury Ferreira and Pedro Botelho and Fernanda Silva and Daniel Casanova and Rafael Faustino and Frederico Oliveira and Arlindo Galvão Filho and Anderson da Silva Soares},
  year      = {2026},
  booktitle = {{Interspeech 2026}},
  pages     = {174--179},
  doi       = {10.21437/Interspeech.2026-2960},
  issn      = {2958-1796}
}

License

MIT. See the CAL-MOS repository for the project license. The W2V-BERT-2.0 backbone remains subject to its own license.