Evaluating Compositional Structure in Audio Representations
This addresses the problem of assessing compositional structure in audio for researchers in machine learning and audio processing, though it is incremental as it adapts existing ideas from vision and language to audio.
The authors tackled the lack of evaluation protocols for compositionality in audio representations by proposing a benchmark with two tasks (A-COAT and A-TRE) and large synthetic datasets, resulting in the first such benchmark for audio embeddings.
We propose a benchmark for evaluating compositionality in audio representations. Audio compositionality refers to representing sound scenes in terms of constituent sources and attributes, and combining them systematically. While central to auditory perception, this property is largely absent from current evaluation protocols. Our framework adapts ideas from vision and language to audio through two tasks: A-COAT, which tests consistency under additive transformations, and A-TRE, which probes reconstructibility from attribute-level primitives. Both tasks are supported by large synthetic datasets with controlled variation in acoustic attributes, providing the first benchmark of compositional structure in audio embeddings.