A Comparative Study of Traditional and Deep Learning Models Using Emotional Customer Data in Automotive Sales Forecasting

This study investigates the effectiveness of traditional and deep learning models in predicting automotive sales outcomes using customer acoustic and emotional parameters extracted from sales call recordings. A primary dataset of sales conversations was processed using an internal tool, which converts audio into numerical data based on 11 acoustic features such as pitch, volume, and speech rate. Traditional machine learning models, including AdaBoost, Support Vector Machine (SVM), and Random Forest, were compared with deep learning models like Dense Neural Networks (DNN) and TabNet. The optimized DNN model achieved the highest performance, with an F1-Score of 0.7298, balancing precision and recall effectively. TabNet provided competitive results with an F1-Score of 0.7237, offering additional benefits of feature interpretability through its attention mechanism. These findings suggest that while traditional models, particularly AdaBoost, perform adequately, deep learning models, especially DNN and TabNet, are better equipped to capture the complex, non-linear relationships in acoustic and emotional data. The results provide actionable insights for optimizing sales strategies through data-driven decision-making. It offers a foundation for future advancements in predictive analytics within automotive industries using voicebased datasets.