In the age of increasing machine-mediated communication, the ability to detect emotional nuances in speech has become a critical competency for intelligent systems. This paper presents a robust Speech Emotion Recognition (SER) framework that integrates a hybrid deep learning architecture with a real-time web-based inference interface. Utilizing the RAVDESS dataset, the proposed pipeline encompasses comprehensive preprocessing, data augmentation techniques, and feature extraction based on Mel-Frequency Cepstral Coefficients (MFCCs), Chroma features, and Mel-spectrograms. A comparative experiment was run against a standard machine learning classifier such as K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest, and XGBoost. The experimental results indicate that the CNN-BiLSTM-Conv1D model proposed is much better as compared to conventional models with a state-of-the-art classification accuracy of 94%. The model was further evaluated using ROC-AUC curves and per-class performance metrics. It was subsequently deployed using a Flask-based web interface that enables users to upload voice inputs and receive real-time emotion predictions. This end-to-end system addresses the shortcomings of earlier SER approaches---such as limited temporal modeling and reduced generalization---and showcases practical applicability in domains like mental health monitoring, virtual assistants, and affective computing.