Enhancing Aphasia Speech Interpretation using Small Language Models (SLMs)

Aphasia is a severe communication disorder that significantly impairs an individual's ability to convey and process language, often resulting from stroke-related damage to brain regions critical for speech and language functions. With the emergence of Large Language Models (LLMs), their potential has been explored in various text-based tasks due to their exceptional language understanding capabilities, which are particularly valuable in medical applications where access to specialised data is crucial yet frequently restricted. In this paper, we present our research on leveraging Tiny and Small Language Models (SLMs) to improve speech interpretation for people living with aphasia (PwA). Through benchmarking several LLMs, we established performance benchmarks to guide the development of our SLM-based solution. Our findings indicate that chain-of-thought prompting significantly enhances interpretation accuracy (median similarity score: 0.68 vs. 0.64 for zero-shot), with larger SLMs (e.g., Phi4-mini:3.8b) outperforming smaller counterparts while maintaining clinical utility. Notably, compact models like Qwen2.5:1.5b achieved competitive results, demonstrating feasibility for re-source-constrained settings. This work advances accessible, privacy-preserving assistive technology for aphasia, balancing computational efficiency with clinical relevance.