UTMOST

Official implementation of UTMOST: A Multi-Task Model for Assessing Emotional Speech Synthesis.

UTMOST is a single model that returns quality MOS (qMOS), emotion-similarity MOS (eMOS), valence, arousal, dominance, and the perceived emotion category of the input speech.

Model weights: coming soon. The code is available now. The pretrained weights were trained on the VoiceMOS Challenge 2026 Track 2 dataset, whose rights are held by the challenge organizers, so we can release them only after the organizers make the dataset public. We will publish them on the Hugging Face Hub (sarulab-speech/UTMOST) as soon as that happens. Until then, utmost.load() cannot download them. T02, our Track 2 submission to the VoiceMOS Challenge 2026, will also be released in this repository.

Outputs

Key	Output
qmos	Quality MOS (1–5)
emos	Emotion-similarity MOS: how closely the utterance matches the requested emotion (1–5)
valence, arousal, dominance	Valence, arousal, and dominance (1–5)
category	Probability that a listener selects each perceived emotion category (angry, happy, neutral, sad, surprise)

The qMOS and eMOS heads predict listener-dependent scores. At inference, UTMOST averages its predictions over 64 listener embeddings, selected at approximately equal intervals from the embeddings of the 262 listeners in its training data.

Model

UTMOST fine-tunes one WavLM-Large model jointly on all six outputs and augments its shared acoustic representation with a fixed 1,025-dimensional side feature extracted by a frozen UTMOS22-Strong model. The eMOS branch is additionally conditioned on the requested emotion and on predicted emotion-category features. UTMOST was trained on all training ratings of the VoiceMOS Challenge 2026 Track 2 (12,746 utterances and 262 listeners) and has 321.9M trainable parameters.

Results on the VoiceMOS Challenge 2026 Track 2 evaluation set, as reported in the paper. Scores are utterance-level SRCC, except CAT error, the mean absolute difference between predicted category probabilities and listener selection proportions (lower is better).

qMOS	eMOS	Valence	Arousal	Dominance	CAT error
0.783	0.752	0.735	0.774	0.696	0.153

Installation

pip install git+https://github.com/sarulab-speech/UTMOST

Python 3.9 or later and PyTorch 2.0 or later are required. On Windows, pip installs the CPU-only build of PyTorch; to use an NVIDIA GPU, install a CUDA build first, for example pip install torch --index-url https://download.pytorch.org/whl/cu130.

The model weights are downloaded from the Hugging Face Hub (sarulab-speech/UTMOST) on first use. If the repository requires access, log in first with hf auth login, or with from huggingface_hub import login; login() in Python.

Usage

Python

import utmost

model = utmost.load()  # uses a GPU if one is available
out = model.predict("sample.wav", emotion="Happy")
print(out["qmos"], out["emos"], out["category"])

# Several files, one requested emotion per file
outs = model.predict(["a.wav", "b.wav"], emotion=["Sad", "Angry"])

# A waveform in memory
out = model.predict(waveform, sample_rate=24000, emotion="Neutral")

emotion is the emotion the synthesizer was asked to express: Angry, Happy, Neutral, Sad, or Surprise. If it is omitted, UTMOST uses the category with the highest predicted probability and reports emotion_source: "predicted" in the output.

With all_emotions=True, the output also holds eMOS, valence, arousal, and dominance for each of the five emotions under by_emotion.

utmost.load() runs in half precision on a GPU and in single precision on a CPU. Use utmost.load(fp16=False) for single precision on a GPU.

torch.hub

import torch

model = torch.hub.load("sarulab-speech/UTMOST", "utmost", trust_repo=True)
model.predict("sample.wav", emotion="Happy")

Command line

utmost wavs/ --emotion happy -o scores.csv
utmost wavs/ --emotion-table emotions.txt -o scores.csv  # one "file.wav|Sad" per line

Citation

@misc{utmost,
  title  = {{UTMOST}: A Multi-Task Model for Assessing Emotional Speech Synthesis},
  author = {Park, Joonyong and Kim, Hounsu and Park, Yoonjeong and Kawamatsu, Ryota and
            Nakata, Wataru and Saito, Yuki and Fukayama, Satoru and Nam, Juhan},
  note   = {Under review}
}

License

TBD (not yet decided). Third-party code and weights included in UTMOST keep their own licenses; see THIRD_PARTY_NOTICES.md.