Hugging Face's logo

alefiury
/
CALMOS-XLS-R-300m-SingMOS

Feature Extraction
Transformers
Safetensors
calmos
audio
speech
speech-quality-assessment
mos-prediction
speech-foundation-models
cal-mos
interspeech-2026
custom_code
Model card Files
xet
 Community
CALMOS-XLS-R-300m-SingMOS
File size: 2,685 Bytes
3f3c53d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
---
license: mit
library_name: transformers
tags:
- audio
- speech
- speech-quality-assessment
- mos-prediction
- speech-foundation-models
- cal-mos
- interspeech-2026
---

# CAL-MOS 路 XLS R 300m 路 SingMOS

**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 **XLS R 300m**, trained on **SingMOS**.

[Paper](https://www.isca-archive.org/interspeech_2026/ferreira26_interspeech.html) 路 [arXiv](https://arxiv.org/abs/2609.14956) 路 [GitHub](https://github.com/Akcit-MOS-Prediction/CAL-MOS)

## Model

| | |
|---|---|
| **Backbone** | [`facebook/wav2vec2-xls-r-300m`](https://huggingface.co/facebook/wav2vec2-xls-r-300m) |
| **Dataset** | `SingMOS` |
| **Strategy** | Adapters + Mean (A+M) |
| **Backbone** | Frozen |
| **Pooling** | Mean |
| **Input sampling rate** | 16,000 Hz |

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.

```python
from transformers import AutoModel

model = AutoModel.from_pretrained(
    "alefiury/CALMOS-XLS-R-300m-SingMOS",
    trust_remote_code=True,
).to("cuda")

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

Batch inference is also supported:

```python
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

```bibtex
@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 Galv茫o 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](https://github.com/Akcit-MOS-Prediction/CAL-MOS) for the project license.
The `XLS-R-300m` backbone remains subject to its own license.