Speech models get fixes for trouble understanding updated text knowledge
Repairing Speech Access After Text Updates with Frozen Affine Fields
Computation and Language
Summary
Speech-based AI systems learn a lot from written text, but sometimes when the written text is updated, the speech systems don’t understand the new info well. The authors propose a way to fix this by keeping the main model frozen and adding a small operator that adjusts how speech queries access updated knowledge. They test their method, called AFFIRM, and find it can improve speech comprehension in some cases compared to other approaches. This shows that updating text knowledge and speech access needs separate attention and fixing.
What this means in practice
- •For voice assistant developers: Fix speech-based AI systems so they better understand information after text updates without retraining the whole model.
- •For speech recognition engineers: Improve speech query handling in multi-modal language models by applying detachable correction operators to frozen speech states.
Authors
Ming-Hao Hsu, Yuxiang Wang, Yingda Shen, Zhizheng Wu
Abstract
Speech-capable language models inherit most of their knowledge from text-rich language backbones, and continued model maintenance is likewise dominated by text updates. Knowledge acquired through a text update can nevertheless remain difficult to access when the same query is spoken. We formulate this mismatch as post-update speech access repair, where the knowledge-bearing checkpoint stays frozen and a detachable operator repairs speech-only access in one model pass. We introduce Affine Fields for Interface Repair across Modalities (AFFIRM), which first learns and freezes a base correction for speech hidden states, then adds query-specific corrections with zero mean over the operator-training queries. Paired speech and text states provide the coordinates that condition these corrections on each query. On balanced synthetic-speech PopQA, paired comparisons do not establish a clear advantage for AFFIRM over global SALSA steering or low-rank ReFT interventions under the same frozen single-state constraint. On recorded SLUE-SQA-5 questions with text passages, a full-scale PopQA-fitted AFFIRM operator transfers without refitting and improves ROUGE-L by 0.0941 over its paired No Repair baseline. Low-rank ReFT interventions also transfer positively, whereas the matched SALSA variant falls 0.0707 below its paired baseline. These results show that text acquisition and speech access can diverge after an update and should be verified and repaired separately.