Voice Outperforms Text in Identifying Personality Traits in Bangla: A Comparative Study

In recent years, the exploration of personality traits through both text and speech has garnered significant attention, offering valuable insights into psychological profiling, human-computer interaction, and personalized services. Previous studies have primarily focused on identifying personality traits from user-written text. Although some efforts have been made to detect personality traits using Bangla text, the results have been largely unsatisfactory. Despite the fact that there is a well-established relationship between speech and personality traits, there is no research on automatically determining personality traits from Bangla speech. To address this gap, we have conducted an in-depth study that highlights the effectiveness of Bangla voice-based models. Our findings have revealed that these models outperform text-based models in accurately detecting personality traits, showing significant potential in this domain. We compiled a substantial dataset of Bangla speech recordings, covering diverse demographic backgrounds and personality traits. Utilizing state-of-the-art deep learning models, we proposed two advanced systems for personality trait detection from Bangla speech, specifically employing VGG19 and ResNet50 architectures. Our deep learning framework incorporated convolutional neural networks to capture temporal dependencies and spectral features inherent in speech. Through extensive experimentation, we evaluated the efficacy of our approach on the speech dataset and compared it to text-based methods presented in [1]. Our results demonstrated promising outcomes, confirming the practicality of using voice for personality trait detection in Bangla. Notably, the VGG19 model emerged as the top performer, achieving an accuracy score of 79.17%, which is much higher than text based methods.