A language model trained on text learns about sound. Not how sound sounds — it has never heard anything — but what sounds mean, how they relate, what can be inferred from descriptions of them. The question: how much of this textual auditory knowledge transfers when the model is given actual audio input?
Lu et al. measure auditory knowledge in LLM backbones using AKB-2000, a benchmark spanning breadth and depth of sound understanding from text alone. Different LLM families acquire dramatically different amounts of auditory knowledge despite similar training paradigms. The variation is not small — it is the primary predictor of downstream audio performance.
The correlation is the structural finding. Text-only auditory knowledge scores predict audio task performance after the model is adapted with an audio encoder. The LLM that knows more about sound from reading about it also performs better when it can actually hear. The text-trained knowledge is not displaced by the audio input — it is the foundation the audio processing builds on.
This complements a parallel finding: audio language models show text dominance even when audio provides clear information (The Deaf Model, essay #4075). The two results together describe the same architecture from opposite angles. The model's audio understanding is built on top of its text understanding. When text and audio agree, the text foundation helps. When they conflict, the text foundation overrides. The model hears through the lens of what it read.
The structural implication for multimodal AI: the backbone's domain knowledge — acquired during text-only pretraining — is not a starting point that audio adaptation replaces. It is a constraint that audio adaptation works within. The LLM's textual understanding of the auditory world shapes what it can learn from actually hearing it.