Abstract
PURPOSE: Development of aphasia therapies is limited by clinician shortages, patient recruitment challenges, and funding constraints. To address these barriers, we introduce Agent-Based Conversational Dialogue (ABCD), a novel method for simulating goal-driven natural spoken dialogues between two conversational artificial intelligence (AI) agents-AI clinician (Re-Agent) and AI patient (AI-Aphasic), which vocally mimics aphasic errors. Using ABCD, we simulated response elaboration training between both agents with stimuli varying in semantic constraint (high via pictures, low via topics). Rather than resource-intensive fine-tuning, we leveraged prompt engineering, chain-of-thought (CoT) and zero-shot techniques for rapid, cost-effective agent development, and piloting. METHOD: Built on OpenAI's GPT-4o as the foundational large language model, Re-Agent and AI-Aphasic were supplemented with external speech-to-text and naturalistic text-to-speech application programming interfaces to create a multiturn, dynamic dialogue system in English. We used it to evaluate Re-Agent's conversational performance across four experimental conditions (CoT + picture, CoT + topic, zero-shot + picture, zero-shot + topic) and aphasic error at two levels: word and discourse errors. Re-Agent's performance was measured using three discourse metrics: global coherence, local coherence, and grammaticality of utterances. RESULTS: Overall, Re-Agent performed accurately in all the discourse metrics across all levels of semantic parameter, prompting technique and aphasic error. The results also indicated that well-crafted zero-shot prompts induce more direct and logically related responses that are robust to adversarial aphasic speech inputs, whereas CoT might lead to responses that slightly lose local coherence due to additional complex reasoning chains. CONCLUSIONS: ABCD represents a foundational computational approach to accelerate the innovation and preclinical testing of conversational AI partners for speech-language therapy. ABCD circumvents the barriers of collecting diverse errorful speech samples for clinical conversational AI fine-tuning. As AI systems-including large language models and speech technologies-advance rapidly, ABCD will scale accordingly, further enhancing its potential for clinical integration.