Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The T12 System for AudioMOS Challenge 2025: Audio Aesthetics Score Prediction System Using KAN- and VERSA-based Models

Katsuhiko Yamamoto, Koichi Miyazaki, Shogo Seki

TL;DR

The paper proposes AESCA, an AES prediction system for AMC25 Track 2 that fuses a KAN-based audiobox aesthetics predictor with a VERSA-based regression model. It leverages iterative pseudo-labeling and unlabeled data (VMC22) to improve generalization, and combines four KAN models with a VERSA model via stacking. Experimental results show superior correlation metrics across utterance- and system-level axes, indicating the efficacy of ensembling diverse, metric-informed predictors. The work demonstrates how learnable KAN activations and a broad metric ensemble can better capture audio aesthetics in synthetic speech and music generation contexts. Overall, AESCA advances reference-free AES prediction by integrating semi-supervised learning and ensemble techniques to align predictions with human judgments.

Abstract

We propose an audio aesthetics score (AES) prediction system by CyberAgent (AESCA) for AudioMOS Challenge 2025 (AMC25) Track 2. The AESCA comprises a Kolmogorov--Arnold Network (KAN)-based audiobox aesthetics and a predictor from the metric scores using the VERSA toolkit. In the KAN-based predictor, we replaced each multi-layer perceptron layer in the baseline model with a group-rational KAN and trained the model with labeled and pseudo-labeled audio samples. The VERSA-based predictor was designed as a regression model using extreme gradient boosting, incorporating outputs from existing metrics. Both the KAN- and VERSA-based models predicted the AES, including the four evaluation axes. The final AES values were calculated using an ensemble model that combined four KAN-based models and a VERSA-based model. Our proposed T12 system yielded the best correlations among the submitted systems, in three axes at the utterance level, two axes at the system level, and the overall average.

The T12 System for AudioMOS Challenge 2025: Audio Aesthetics Score Prediction System Using KAN- and VERSA-based Models

TL;DR

The paper proposes AESCA, an AES prediction system for AMC25 Track 2 that fuses a KAN-based audiobox aesthetics predictor with a VERSA-based regression model. It leverages iterative pseudo-labeling and unlabeled data (VMC22) to improve generalization, and combines four KAN models with a VERSA model via stacking. Experimental results show superior correlation metrics across utterance- and system-level axes, indicating the efficacy of ensembling diverse, metric-informed predictors. The work demonstrates how learnable KAN activations and a broad metric ensemble can better capture audio aesthetics in synthetic speech and music generation contexts. Overall, AESCA advances reference-free AES prediction by integrating semi-supervised learning and ensemble techniques to align predictions with human judgments.

Abstract

We propose an audio aesthetics score (AES) prediction system by CyberAgent (AESCA) for AudioMOS Challenge 2025 (AMC25) Track 2. The AESCA comprises a Kolmogorov--Arnold Network (KAN)-based audiobox aesthetics and a predictor from the metric scores using the VERSA toolkit. In the KAN-based predictor, we replaced each multi-layer perceptron layer in the baseline model with a group-rational KAN and trained the model with labeled and pseudo-labeled audio samples. The VERSA-based predictor was designed as a regression model using extreme gradient boosting, incorporating outputs from existing metrics. Both the KAN- and VERSA-based models predicted the AES, including the four evaluation axes. The final AES values were calculated using an ensemble model that combined four KAN-based models and a VERSA-based model. Our proposed T12 system yielded the best correlations among the submitted systems, in three axes at the utterance level, two axes at the system level, and the overall average.

Paper Structure

This paper contains 12 sections, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. Submitted T12 System (AESCA)
  3. Baseline Model
  4. KAN-based Audiobox Aesthetics
  5. VERSA-based predictor
  6. Ensemble model
  7. Experiments
  8. Dataset
  9. Experimental Conditions
  10. Training Strategies
  11. Results
  12. Conclusion

Figures (1)

  • Figure 1: Overview of the proposed T12 system for AMC25 Track 2. The system is a weighted ensemble model consisting of a KAN-based Audiobox-Aesthetics predictor (a) and a VERSA-based regression predictor (b).