When Voice AI Has the Evidence but Follows the Text

TL;DR for operators

Audio AI can retain information about a speaker’s voice and still produce an answer that follows a contradictory written description. In controlled experiments, audio language models favored conflicting textual claims at a reported mean rate of 78%, rising to 86% when sarcasm and pacing were excluded.

The research traces this failure to how textual information influences downstream representations, rather than simply to missing acoustic information. A targeted attention intervention substantially reduced text dominance during free-form generation in one model, but the benefit depended on intervention placement and evaluation format.

For contact-center and voice-assistant teams, the immediate lesson concerns diagnosis and testing. Transcription quality and ordinary classification accuracy cannot establish whether a system uses vocal evidence appropriately when sources disagree. Acoustic capability, conflict handling, free-form responses, and preservation of valid behavior need separate tests. The proposed intervention is a research result, not a deployment-ready fix.

A voice assistant can follow the wrong evidence

Imagine a customer speaking with audible frustration while a written description supplied to an AI assistant says the customer sounds happy. The assistant is asked to identify the customer’s emotional tone and answers according to the written description.

That failure might appear to show that the system cannot recognize frustration in speech. But it could also mean the system extracted the relevant vocal information and subsequently gave greater weight to the written assertion.

Ankur Banerjee and coauthors investigate precisely this distinction in Text Dominance in Paralinguistic Conflict: Tracing Suppression Through Audio LLM Circuits.¹

Their study constructs a 3,594-item benchmark spanning emotional prosody, speaker-category judgments, and sarcasm or pacing. Across six audio language models, the reported mean text-dominance rate on modality-split scored conflicting inputs is 0.78, or 0.86 excluding sarcasm/pacing. The mean excludes an earlier SALMONN run that lacked modality-split scoring, although SALMONN remains in the six-model appendix comparisons.

The results establish a recurring behavioral problem across different encoder families. They do not establish that every model fails for the same internal reason.

The acoustic evidence survives, but its influence weakens

The researchers examined Qwen2-Audio-7B-Instruct more closely, first checking whether the model could perform the relevant acoustic task without misleading textual claims.

They then used linear probing: training a simple classifier to recover speaker-category information from frozen internal representations.

Speaker-gender classification reached 1.000 accuracy from audio-encoder layer 13 onward, and the information remained recoverable after projection into the language model’s input space.

The result is specific to the tested speaker-category task. It does not prove perfect perception of emotion or sarcasm, nor does voice-based category classification establish someone’s actual gender identity.

Nevertheless, it demonstrates that relevant information can be present internally even when the final answer follows contradictory text.

A second analysis estimated how much different model components contributed to the decision. Attribution patching assigned contribution magnitudes of 3.29 to text and 0.18 to audio, approximately an 18:1 ratio. Acoustic attribution also became very small in later language-model layers.

These attribution magnitudes are specific to the studied model and procedure, not universal modality weights. Together with probing, however, they shift the diagnosis toward downstream information use.

Making the audio stronger did not restore the decision

If acoustic information is already represented, increasing its strength might seem a reasonable repair.

The researchers tested activation back-patching, which transfers internal representations between model layers, and direct amplification of audio representations. Neither reliably restored audio-grounded decisions.

At fivefold amplification, measured capability fell from 1.000 to 0.100 without the intended repair. Increasing back-patching destination-window width also failed to resolve the conflict.

The back-patching result requires an additional qualification. The study replicated a previously proposed lexical-conflict intervention on 300 resynthesized examples. It repaired 14 predictions but damaged 12, producing a pooled accuracy gain of only 0.007 (95% paired confidence interval: -0.027 to 0.040; McNemar p = 0.845).

Because the procedure did not produce a measurable gain even on its reconstructed source task, the researchers cannot conclude that lexical and paralinguistic conflicts necessarily involve different mechanisms.

For model engineers, this is a distinction between an unsuccessful intervention and a validated explanation for why it failed.

The conflicting claim can influence intermediate tokens

The more consequential experiment examined how a false textual claim affects the model’s eventual answer.

A language model produces representations for successive tokens. The original claim can influence intermediate tokens, including question and answer-option representations, which then carry that information toward the final prediction.

The researchers tested these pathways using attention knockout, selectively preventing tokens from attending to the conflicting claim.

Blocking only the final answer position’s direct access to that claim produced essentially no improvement. Blocking access across subsequent token positions changed the outcome substantially.

In the broad routing experiment, audio-consistent prosody accuracy increased from 0.000 to 0.533, while speaker-category accuracy rose from 0.067 to 0.717.

But the intervention also reduced measured capability on otherwise valid inputs: from 1.000 to 0.367 for prosody and to 0.800 for speaker identity.

The contrast supports a proposed intermediate token-relay mechanism. Once conflicting textual information has affected intervening representations, removing one direct attention connection at the answer is insufficient.

The experiment demonstrates causal effects of particular attention restrictions in the tested model. It does not establish a complete, universally applicable circuit.

A narrower intervention works during generation

The researchers next examined free-form responses rather than only selecting from predefined answer choices.

They measured generated text-dominance rate (gen-TDR): how often generated descriptions mentioned the false textual label.

Mid-layer attention knockout produced the following results in Qwen2-Audio-7B-Instruct:

Source: Section 4.6, Table 3. Results describe the tested free-form generation configurations and relatively small experimental subsets.

The intervention therefore reduced false-label mentions by 62.5 percentage points for prosody and 56.7 points for speaker category, without degrading the reported matched-item or text-only accuracy.

Stronger suppression was not necessarily preferable. Blocking all layers reduced prosody gen-TDR further to 0.182 but lowered matched-item accuracy to 0.667.

Crucially, the effective mid-layer configuration did not yield a corresponding improvement under ordinary closed-set multiple-choice evaluation. Question and option tokens can provide additional pathways for textual information.

This means an evaluation that measures only selected answers may miss a change that appears during autoregressive generation. Conversely, a lower gen-TDR is not sufficient evidence that every generated description is correct.

What voice-AI teams should measure differently

For a contact-center team selecting an emotion-aware voice assistant, the research suggests a more discriminating evaluation sequence.

Start with audio-only examples to establish whether the system can identify the target vocal cue at all. Then test matched textual descriptions, deliberately contradictory descriptions, and free-form responses using the same acoustic evidence.

The distinction guides remediation. If audio-only performance is weak, improving acoustic capability or choosing another model may be appropriate. If audio-only performance is strong but contradictory text changes the answer, investigating how information is combined and routed becomes more relevant.

Any attempted repair must also be tested on ordinary valid interactions, reliable textual evidence, and unconstrained generated responses. A reduction in conflict-specific errors can lose business value if it simultaneously reduces correct judgments elsewhere.

Targeted attention intervention offers a possible engineering research direction, particularly where model-weight changes are undesirable. The paper does not demonstrate production reliability, latency feasibility, or financial returns from such interventions.

The boundaries of the finding

The speech sources include acted emotional examples, the constructed conflict prompts are in English, and detailed mechanistic experiments concentrate on one model. Sarcasm/pacing performance is too weak under audio-only testing to support the same suppression interpretation.

The study also does not resolve what happens when acoustic evidence is ambiguous but the textual assertion is reliable. Unconditional suppression could then remove information the system ought to use.

Operational use additionally requires validation on natural conversations, appropriate privacy and consent protections for vocal-attribute inference, and licensing checks before commercial reuse of research materials.

The decision changes before the model does

The paper’s strongest contribution is its explanation of how an audio model can possess relevant evidence without allowing that evidence to determine its answer.

For operators, this changes the first question after a failure. The issue may be acoustic perception, downstream evidence use, or an evaluation format that conceals the relevant behavior.

Those possibilities require different tests and different investments. Treating them separately provides a more defensible basis for improving voice AI than assuming a stronger audio signal will automatically produce a better judgment.

Cognaptus: Automate the Present, Incubate the Future.