Large Language Models (LLMs) have taken the world by storm since their introduction a few years ago. While these tools excel in mimicking human behavior and show promise for speech analysis, they were primarily trained on text rather than speech. This creates significant gaps between their current capabilities and the demands of speech-based applications. Fine-tuning LLMs on speech data is a critical step forward, requiring transcribed datasets annotated by skilled linguists.

To bridge this gap, we are engaged in several collaborative projects with NLP labs, focusing on speech annotation in both English and Hebrew. These efforts aim to adapt LLMs to handle spontaneous speech more effectively and to enhance our ability to study this complex and natural form of communication. Additionally, we utilize LLM tools to analyze brain responses to spontaneous speech, exploring neural processing at various levels of language representation.