Reproduce “Hearing the Unspoken: Language Model Priors for Acoustic Adversarial Attacks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiani Xie, Andrew C. Cullen, Paul Montague, Benjamin I. P. Rubinstein
Semantic Gambit (SG) attacks live speech recognition by using an LLM to predict likely upcoming words from a speech prefix, then conditioning an acoustic perturbation generator on the prediction and prefix audio. On Wav2Vec 2.0 Large and LibriSpeech test-clean, SG reaches 35.63% corpus-level WER versus 15.15% for an acoustic-only ablation at zero delay; average measured pipeline latency is 663 ms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.