KOMOS

Automatic Mean Opinion Score Estimation for Korean Speech

KOMOS is a Korean mean opinion score predictor for overall speech quality across read and spontaneous speech. It combines a fully fine-tuned Whisper-medium encoder with a listener-conditioned prediction head and returns a score on the 1–5 MOS scale.

Highlights

• Trained with 5,000 ratings from 29 native Korean listeners for 1,000 utterances.
• Covers real recordings, synthesized speech, and degraded speech.
• Supports PyTorch and ONNX Runtime inference.

Results

Human MOS alignment

On the Korean test set, KOMOS follows the human MOS scale more closely than UTMOS v1.

Mean predicted MOS by human-MOS bucket after validation-set affine calibration. The dashed line shows ideal alignment.

Overall test performance

Utterance-level correlation with human MOS on the 100 utterance test set. UTMOS v1 and v2 are evaluated as English-trained zero-shot baselines, while KOMOS is fine-tuned on Korean quality ratings.

Speaker-disjoint and text-disjoint evaluation

Utterance-level SRCC on test subsets whose speakers or texts are absent from both the training and validation sets.

Installation

Python 3.10 or later is required.

Install KOMOS from PyPI:

pip install komos

To work from a local checkout instead:

git clone https://github.com/upskyy/komos.git
cd komos
uv sync --no-dev

Quickstart

PyTorch

from komos import KOMOS

model = KOMOS.from_pretrained("upskyy/komos").eval()
score = model.score("audio.wav")
print(score)

Move the model to CUDA with model.to("cuda") when GPU inference is needed.

torch.hub

import torch

model = torch.hub.load("upskyy/komos:v0.0.1", "komos", trust_repo=True)
print(model.score("audio.wav"))

ONNX Runtime

pip install "komos[onnx]"

from komos.onnx import KOMOSOnnx

model = KOMOSOnnx.from_pretrained("upskyy/komos")
score = model.score("audio.wav")
print(score)

Inference

score() accepts an audio path or a mono NumPy array or PyTorch tensor with an explicit sample rate. File inputs are downmixed to mono and resampled to 16 kHz.

• score() evaluates up to the first 10 seconds and returns a validation-calibrated score clipped to the 1–5 range.
• score_long_form() supports recordings of up to 60 seconds and computes a duration-weighted mean of raw scores from non-overlapping 10-second chunks before calibration.
• Inputs longer than 60 seconds should be processed in external batches.

Citation

If you use KOMOS in your research, please cite:

@misc{ha2026komos,
  title         = {KOMOS: Automatic Mean Opinion Score Estimation for Korean Speech},
  author        = {Sangchun Ha and Junwon Moon and Kyuhong Shim},
  year          = {2026},
  eprint        = {2610.15353},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  url           = {https://arxiv.org/abs/2610.15353},
}

License

KOMOS is released under the MIT License. See LICENSE and NOTICE.