APEX: Audio Prototype EXplanations for Classification Tasks

Audio explanations need domain-specific approaches: The paper demonstrates that applying vision-based explanation techniques to audio spectrograms is insufficient because visual and acoustic signals have fundamentally different properties. APEX addresses this by offering four distinct explanation perspectives (square-based, time-based, frequency-based, and time-frequency-based) tailored to how sound actually works.

Prototype-based explanations offer intuitive understanding: Rather than using abstract gradient-based methods, APEX explains audio classifier decisions by showing actual audio examples (prototypes) that influenced the decision. The paper claims this provides "greater semantic clarity than standard gradient-based methods," making it easier for non-experts to understand why a system made a particular classification.

The framework preserves model integrity: APEX operates as a "post-hoc" system that requires no modification or fine-tuning of the original audio classifier and "strictly preserves output invariance." This means organizations can add explainability to existing audio AI systems without altering their performance or requiring retraining.

Multi-dimensional acoustic similarity requires multiple explanations: The paper recognizes that sound similarity is complex—it can involve sudden events (transient events), patterns over time, specific frequency ranges, or combinations thereof. By offering four complementary perspectives, APEX captures this multidimensional nature rather than forcing audio into a single explanation framework.