Stacking Complementary CLAP Embeddings for Improving Text-Audio Alignment Correspondence Scoring

Text-audio alignment correspondence scoring assesses whether an audio clip satisfies a natural-language prompt by combining semantic grounding, acoustic event recognition, and perceptual quality. We study this problem and compare a wide range of frozen representations: BEATs, three LAION CLAP variants, MS-CLAP, WavLM, AudioMAE, EAT, Whisper encoder features, and wav2vec2 CTC features. A clear pattern emerges: audio-only self-supervised learning (SSL) and speech recognition (ASR) features can improve calibration metrics but do not consistently improve the challenge-ranking metric. By contrast, contrastive audio-language models make complementary errors. We therefore propose CLAP stacking, a lightweight method that combines multiple CLAP embedding spaces. Our final frozen-feature system, evaluated on the scored XACLE test set, improves SRCC from 0.5521 for the best single MS-CLAP pipeline to 0.5934 and reduces MSE from 3.2500 to 2.9418. This performance would fall between official ranks 6 and 7, far above the official baseline SRCC of 0.3345.