STAG identifies audio clues behind spoken captions in AI models
What Did the MLLM Hear? Token-Level Spectro-Temporal Grounding for Audio MLLM Explainability
SoundComputer Vision and Pattern RecognitionMachine Learning
Summary
Audio-based AI models can describe sounds in detail, but it's unclear which exact parts of a sound led to each word in the description. The authors created STAG, a tool that shows which time and frequency parts of an audio clip support each generated word. STAG combines clues from when sounds happen and which pitch ranges are important, helping explain the AI reasoning. Tests show STAG outperforms other explanation methods and proves that removing the identified sound parts reduces the AI’s confidence in its word choices.
What this means in practice
- •For audio software developers: Integrate STAG to highlight which parts of audio inputs influence AI-generated captions in audio understanding apps.
- •For audio surveillance teams: Use STAG to interpret and verify which sound events in complex audio trigger alerts from AI-based monitoring systems.
Authors
Lucia Cascone, Valeria Fraenza, Michele Nappi, Fabio Narducci, Benedetto Simone
Abstract
Audio-based Multimodal Large Language Models (MLLMs) can generate detailed natural-language descriptions of complex acoustic scenes, yet it remains unclear which parts of the input audio support each generated token. This is particularly challenging because acoustic evidence is distributed across time and frequency, and concurrent sound events may overlap temporally while occupying different spectral regions. We introduce STAG, to our knowledge the first post-hoc framework for token-level spectro-temporal grounding of captions generated by audio-based MLLMs. STAG estimates the temporal support for each generated token using target-token-specific vocabulary projections of the encoded audio representations, measures frequency-band relevance through controlled spectral occlusion, and combines the two signals into a spectro-temporal relevance map. We evaluate STAG against ten post-hoc explanation methods across four grounding benchmarks, where it achieves the best event-localization performance on every dataset, and apply it to eight audio-language backbones without parameter updates. Counterfactual deletion further shows that removing the identified evidence selectively reduces confidence in the corresponding event and frequently removes it from the regenerated caption. These results provide behavioral support for the faithfulness and selectivity of the explanations.