Research Summary:
This thesis explores the development of a privacy-focused, conversational system for the QTrobot platform, leveraging a Jetson Orin Nano and free and open-source software (FOSS). Building on prior research that highlights users’ preference for conversational humanoid robots, the system integrates automatic speech recognition (ASR), a local large language model (LLM), and a text-to-speech engine to enable real-time interactions while prioritizing user privacy and transparency. The work also investigates fine-tuning LLMs using synthetic data generated from multiple open-source models, revealing both potential benefits and current limitations. Based on these findings, the thesis proposes a new accuracy metric for evaluating synthetic data quality, aiming to support the creation of larger, more effective datasets. Designed with modularity in mind, the conversation system is adaptable to other robotic platforms, though additional validation is needed to confirm its broader applicability.
Reference:
https://urn.fi/URN:NBN:fi:amk-2024122037775