Prediction of System Usability through Neural NetworksApplied to Conversational User Interfaces

This study investigates the use of Deep Neural Networks (DNN) to predict perceived usability in conversational user interfaces (CUIs), focusing on two systems (ChatGPT and Gemini) and two interaction modalities (text and voice). A structured methodologyinvolving data segmentation, regularization techniques, and crossvalidation was applied to develop predictive models that incorporate demographic variables and user experience. Results highlight the superior performance of the Gemini voice model (RMSE: 0.18, R2: 0.80), followed by ChatGPT in text mode (RMSE: 0.2, R2: 0.78). Text-based interaction with Gemini showed lower predictive accuracy, suggesting underfitting and the need for architectural adjustments. In general, voice-based models demonstrated greater consistency and predictive power, possibly due to a more intuitive user experience. These findings confirm the potential of DNNs to model complex user perceptions and support user-centered design. The proposed approach offers a valuable tool for anticipating usability outcomes and enhancing the personalization of CUI experiences.