Binh Nguyen

1. Reasoning in audio-language models

   Does a model’s explanation reflect what it actually heard?

   Audio-language models can now explain their verdicts, but an explanation may not support the verdict, or may rationalize a wrong one. I build diagnostics for how this reasoning shifts under acoustic and linguistic attacks, and explore reasoning-centric post-training (GRPO) that grounds a model’s answers in acoustic evidence.

   • SARA · EMNLP Findings 2026

2. Robustness of audio deepfake detection

   What does a detector hear when only the words change?

   Anti-spoofing systems are usually stress-tested with acoustic noise. We showed that rewriting the transcript alone can break them: attack success exceeds 60% on several open-source detectors, and one commercial detector falls from 100% to 32% accuracy. I also study how people perceive the same linguistic cues.

   • What You Read Isn’t What You Hear · EMNLP 2025 Oral
   • Turing’s Echo · Interspeech 2025

3. Modality disentanglement and steering

   Can we tell what a model heard apart from what it read?

   Multimodal models can lean on what they read over what they hear. I am working on separating acoustic from linguistic information inside these models and steering them at inference time, so their decisions rest on the right evidence.

   • Ongoing work

Earlier, I worked on graph-based and capsule models for low-resource neural machine translation (COLING 2022, TALLIP 2023, JIFS 2022).