MoEScore: Mixture-of-Experts-Based Text-Audio Relevance Score Prediction for Text-to-Audio System Evaluation
Bochao Sun, Yang Xiao, Han Yin
TL;DR
This paper introduces MoEScore, an objective TARS predictor for Text-to-Audio systems based on a Mixture-of-Experts architecture enhanced with Sequential Cross-Attention (SeqCoAttn). By ensembling four specialized experts—including three CLAP-based backbones and a SeqCoAttn-equipped expert—the model captures both global semantic alignment and fine-grained temporal correspondence, fused via a gating network trained end-to-end. On the XACLE Challenge 2026 dataset, MoEScore achieves a SRCC of $0.6402$, representing a $30.6\%$ improvement over the baseline, and rankings indicate strong performance across LCC and KTAU with reduced MSE. The approach offers a scalable, objective alternative to subjective human evaluations for text-audio relevance, with potential applicability to broader cross-modal evaluation tasks.
Abstract
Recent advances in generative models have enabled modern Text-to-Audio (TTA) systems to synthesize audio with high perceptual quality. However, TTA systems often struggle to maintain semantic consistency with the input text, leading to mismatches in sound events, temporal tructures, or contextual relationships. Evaluating semantic fidelity in TTA remains a significant challenge. Traditional methods primarily rely on subjective human listening tests, which is time-consuming. To solve this, we propose an objective evaluator based on a Mixture of Experts (MoE) architecture with Sequential Cross-Attention (SeqCoAttn). Our model achieves the first rank in the XACLE Challenge, with an SRCC of 0.6402 (an improvement of 30.6% over the challenge baseline) on the test dataset. Code is available at: https://github.com/S-Orion/MOESCORE.