acceptodds
Under review as a conference paper at ICLR 2027

Hearing the Unspoken: Language Model Priors for Acoustic Adversarial Attacks

Abstract

Automatic Speech Recognition (ASR) systems operating in real-time settings must process acoustic input under strict temporal constraints, where transcription decisions are inherently made on incomplete information. This causal constraint serves as an information bottleneck on attackers, significantly limiting attack performance. Our new Semantic Gambit attack overcomes this causal limitation by augmenting the adversary with predictive context derived from a Large Language Model in real-time. Our experiments show that this form of augmentation can elevate the corpus-level Word Error Rate to 35.6% – a three-fold increase over the current state-of-the-art streaming attack. Ultimately, this work reveals how common, low-latency LLM tooling can be exploited to systematically subvert real-time ASR pipelines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
audio, adversarial attacks, automatic speech recognition

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_43xTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies adversarial speech recognition attacks that observe only an audio prefix. It transcribes the observed speech and uses a language model to forecast subsequent text. A multimodal generator combines acoustic and textual representations to produce an attack waveform. The evaluation compares information regimes and injection delays and studies transfer across datasets and recognition models.

Critical strengths
  • Augmenting prefix-conditioned attacks with pretrained linguistic priors is a meaningful, testable idea.
  • The architecture-matched acoustic-only ablation and the broader set of noise, universal, gradient, oracle, and architectural comparisons are useful.
  • The appendices disclose future-window amplitude normalization and measured pipeline latency. Failure to transfer to Whisper is also reported.
  • The within-CTC and cross-dataset results justify further investigation within the measured
Critical weaknesses
  • The headline WER uses zero delay, whereas the implemented pipeline takes 564–729 ms, averaging approximately 663 ms. Neither zero nor half-second delay is supported by this implementation. The one-second setting remains relevant, but its gains over the architecture-matched acoustic baseline are much smaller for longer prefixes. Prospective serving improvements do not validate present zero-delay results.
  • The final perturbation is not strictly prefix measurable: every reported experiment scales it using the future target window's RMS. Disabling gradients through that scalar does not remove inference-time dependence. The appendix's causal-RMS SNR deviations do not demonstrate equivalent attack efficacy or budget compliance. This is a disclosed limitation, but conflicts with the stronger causality claim.
  • AO removes both observed transcript and forecast, so it cannot isolate anticipation. Prefix-only, shuffled/random-forecast, and matched-capacity controls are missing. The weaker target-transcript oracle does not resolve this confounding.
  • A deterministic function of the prefix does not enlarge its sigma algebra. Pretrained features can help a limited generator exploit available information, but Proposition 1 does not establish a new information channel. Strict filtration inclusion/entropy increase requires a specified observation model, and stored prefix audio is also invariant to later emitted noise.
  • Causal attacker access does not establish a streaming victim. CTC does not itself make an encoder causal. The exact chunking, right context, caching, and hypothesis commitment protocol need documentation; whole-utterance Wav2Vec2/HuBERT evaluation cannot substitute for them. wav2vec-S
  • One training seed, varying evaluation subsets across prefix lengths, finite-optimization PGD, and SNR-only perceptual claims limit interpretation. The paper acknowledges some of these limitations but still makes stronger causal/general conclusions.
Questions for authors

Q1. Given measured pipeline latency of 564–729 ms, how should the zero- and half-second-delay results be interpreted? What WER advantage remains when injection follows actual per-example completion times?

Q2. Can you rerun with prefix-only or fixed calibrated gain and report WER, realized SNR, and budget violations? How much does exact future-window amplitude information affect the conclusions?

Q3. What gain remains over a prefix-transcript-only control? Do shuffled or unrelated length-matched forecasts rule out ordinary text features or capacity effects?

Q4. What encoder causality, right context, chunking, caching, and partial-transcript commitment protocol does the victim use? Is this a causal attacker against an offline recognizer or a fully streaming pipeline?

Q5. What precisely constitutes additional information? With fixed model parameters, a deterministic prefix transformation does not enlarge its sigma algebra; is the intended claim about priors, computational accessibility, or new observations?

Q6. Do the central gains persist across training seeds and a common evaluation subset shared by all prefix configurations?

Reviewer_QPuDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a white-box, real-time adversarial attack on ASR. It uses an LLM to forecast upcoming spoken words. This textual forecast is combined with the most recently spoken audio and its transcript to train a small perturbation generator. The generator produces a perturbation that, when added to the incoming audio, maximizes the loss of the ASR model.

Critical strengths
  1. Using an LLM forecast is a novel way to enhance real-time adversarial attacks.
  2. The prefix × delay grid, the AO ablation (same architecture, no text), and the GT oracle are the right kinds of controls.
Critical weaknesses
  1. The result that SG outperforms GT is counterintuitive. GT has access to the exact target content, while SG only tries to estimate it with an LLM forecast. Yet, surprisingly, SG performs better than GT. How is this possible, and why? The motivation and the results of the paper appear to contradict each other: SG uses an LLM forecast to improve the attack, yet its ideal case (GT) performs worse than SG itself.

  2. SG also outperforms PGD, which is likewise hard to explain. PGD is a white-box attack with full access, so the attacker can make it as strong as needed by tuning hyperparameters such as the number of steps and the step size. It should be able to push WER even above 100%(WER can exceed 100%). This suggests that your baselines may not be strong enough. Even the absolute WER values reported for PGD seem unacceptably low. I suggest evaluating stronger attacks, such as an adapted version of AutoAttack, PGD-1000, or the C&W attack.

  3. The inference pipeline is not entirely clear, particularly for the evaluation sets (experimental setup). Some of this information is probably spread across different parts of the paper, but it is hard to locate and piece together. For example, given a 10-second audio clip with prefix = 2 s and delay = 1 s, which segments do you use? Do you take seconds 2–4 as the prefix and seconds 5–8 as the attack window? Or do you take multiple prefixes per sample and average the results? Is WER computed only on the attack window (e.g., seconds 5–8) or on the whole sample?

  4. You should evaluate the forecasting performance of the LLM using an appropriate metric, such as WER or an embedding-level similarity. This would give a sense of how much the LLM actually helps.

  5. There is no guarantee that the LLM output matches the content of the actual next speech segment. This is especially true in a live setting, where you would need to select a lightweight LLM rather than a strong, high-parameter one. What is the relationship between forecasting accuracy and attack effectiveness? Given the contradictions in your results, it would be helpful to see how the attack performs when the future text is random, or when it is perturbed in a controlled way (e.g., random word replacement to create a forecast with 20% WER).

  6. Your approach is trained on a CTC-based architecture, and you show that it does not transfer to Whisper. Is the attack effective on autoregressive models such as Whisper when the generator is trained directly against them? It is possible that your method only works for CTC-based models.

  7. I could not find the white-noise experiment in Table 3, despite your reference to it.

  8. In ASR@τ, what does “ASR” refer to? You define the ASR model as M. Also, what is τ here? Is it the delay? you did not define them and it's confusing.

  9. The paper focuses on the attack side, but what about defenses? A 20 dB SNR perturbation is large and audible. For instance, could a simple noise-cancellation mechanism completely remove your perturbation? For audibility, you should also report better metrics, such as PESQ or STOI.

Questions for authors

See the weaknesses section!

Reviewer_vAHLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an adversarial attack method that leverages large language model (LLM) priors to enhance real-time automatic speech recognition (ASR) attacks. The authors introduce an LLM-based prediction module that forecasts upcoming linguistic content from the observed speech prefix, addressing the challenge that attackers can only observe previously spoken audio and cannot access the future speech they intend to attack. The acoustic prefix, its transcription, and the predicted text are then jointly fed into a multimodal generator to produce adversarial perturbations targeting future speech windows.

Under a 20 dB SNR constraint, the proposed method increases corpus-level WER on Wav2Vec 2.0 from 2.05% to as high as 35.63%, substantially outperforming the architecture-matched acoustic-only baseline (15.15%) and an adapted Neural Voice Camouflage baseline (12.23%).

The paper further evaluates attack effectiveness across different prefix lengths, injection delays, generator architectures, datasets, and victim models.

Critical strengths

The paper addresses a well-motivated and practically relevant problem. The authors clearly identify the causal information constraints faced by streaming ASR attacks and propose leveraging pretrained language model priors to predict future linguistic content. This idea has some novelty and potential research value.

The experimental results demonstrate substantial improvements over acoustic-only streaming attacks. In particular, the comparison between SG and AO is valuable because both methods share the same generator architecture and training procedure, allowing the effect of textual conditioning to be evaluated more directly.

The experimental evaluation is relatively comprehensive. The authors consider different information availability regimes, speech prefix lengths, perturbation injection delays, generator architecture variants, and datasets. The inclusion of both an adapted prior streaming attack baseline and an architecture-matched acoustic-only baseline strengthens the experimental comparisons.

Critical weaknesses
  1. The independent contribution of LLM-based future prediction is not sufficiently validated.

The central idea of the paper is that predicting upcoming linguistic content can improve the effectiveness of streaming adversarial attacks. However, the AO ablation removes both the observed prefix transcription and the LLM-generated continuation. Therefore, the performance difference between SG and AO only demonstrates the benefit of the complete textual conditioning channel, rather than the independent contribution of future prediction.

Although the authors also introduce a GT Oracle baseline, this experiment replaces the predicted continuation with the victim ASR's transcription of the target window. It therefore does not isolate the respective contributions of the prefix transcription and future prediction.

  1. The strongest attack performance is inconsistent with the measured real-time operating conditions.

The paper emphasizes that the proposed method can attack streaming speech in real time, with its strongest configuration achieving 35.63% WER. However, this result corresponds to zero injection delay, whereas Appendix D reports an average end-to-end pipeline latency of approximately 663 ms, ranging from 564 to 729 ms. Therefore, the current implementation cannot satisfy the zero-delay or 0.5 second injection-delay settings.

Although the system can generate a three-second perturbation faster than the duration of the audio itself, under the 1.0-second delay configuration, the fixed universal adversarial perturbation achieves 17.21% WER, compared with 18.89% for SG. While SG still causes some degradation in recognition performance, this does not necessarily demonstrate its effectiveness under the claimed real-time operating conditions.

  1. The explanation of the prefix-length effect is not sufficiently supported.

The authors attribute the decrease in WER as the observed prefix becomes longer to a victim-side anchoring effect and the diminishing benefits of linguistic prediction. However, Appendix G.4 confirms that different prefix lengths are evaluated on different subsets of utterances, with different duration distributions. Moreover, the accuracy of the LLM's future predictions is not directly measured.

  1. The temporal reliability of predictions and their effectiveness under semantic shifts remain unclear.

If the predicted content differs substantially from the actual target speech in terms of vocabulary or semantics, would the perturbations generated from these predictions remain effective?

Under the t= 1.0 injection-delay configuration, the LLM can only predict future content based on the previously observed prefix. Its predictions must account for uncertainty over approximately one second of unobserved speech before the three seconds target window, as well as uncertainty within the target window itself.

This means that semantic shifts, negation, or topic changes may cause the LLM's predictions to diverge from the actual speech.

Could the authors provide an evaluation of future-text prediction accuracy and compare attack effectiveness before and after semantic shifts? It would also be useful to examine the differences in WER between cases where the LLM predictions are accurate and those where they are incorrect.

Questions for authors
  1. Could the authors include a prefix-transcript-only baseline to isolate the contribution of LLM-based future prediction from that of the observed transcription?

  2. Could the authors report LLM prediction accuracy across different prefix lengths? How does this explain the strongest attack performance occurring with only a 1.0 second prefix?

  3. How sensitive is attack effectiveness to incorrect future predictions, particularly under semantic shifts or topic changes? Could the authors compare WER between accurately and inaccurately predicted continuations?

  4. Given the measured pipeline latency of 564–729 ms in Appendix D, are the τ = 0.0 s and τ = 0.5 s configurations practically achievable? If not, could τ = 1.0 s serve as the primary real-time evaluation setting, with shorter delays explicitly identified as idealized reference conditions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.