Fact · Voice and AI
In a 2024 study, fine-tuned GPT-2 and T5 models were not robust to the errors of a speech recognizer by default, but training them on spoken dialogue data made them more robust.
Checked against its source on October 17, 2026
In the source's own words
“The results show that LLMs are not robust to spoken noise by default, however, fine-tuning/training such models on a proper dataset of spoken TODs can result in a more robust performance.”
Are LLMs Robust for Spoken Dialogues?, Seyed Mahed Mousavi et al., Proceedings of IWSDS 2024 (arXiv preprint 2401.02297), 2024, Abstract.
Open the original- What it measures
- The authors transcribed spoken dialogues with a speech recognizer, characterized its errors, and simulated those errors in a larger set of dialogues before testing the models.
- What it does not show
- The models were small fine-tuned models, not current chatbots, and the test was task-oriented dialogue. The result is about the direction of the effect.