SGPA: Spectrogram-Guided Phonetic Alignment for Feasible Shapley Value Explanations in Multimodal Large Language Models

Explaining the behavior of end-to-end audio language modelsvia Shapley value attribution is intractable under native tok-enization: a typical utterance yields over 150 encoder frames,inflating the coalition space by roughly 1042 relative to text;individual audio frames lack standalone meaning; and tokenboundaries that bisect phonetic transitions introduce maskingartifacts. We introduce Spectrogram-Guided Phonetic Align-ment (SGPA), a four-stage pipeline that combines Connec-tionist Temporal Classification forced alignment with spec-tral boundary refinement to produce acoustically stable, word-aligned audio segments. Controlled diagnostics on LFM2-Audio-1.5B with VoiceBench show that SGPA yields a 43×reduction in model evaluations. Statistical testing confirmsthat SGPA significantly alters attribution concentration whilepreserving the global cumulative profile, establishing it as afeasibility-enabling layer for audio explainability.