PENGEMBANGAN SPEECH EMOTION RECOGNITION BERBAHASA INDONESIA DENGAN EKSTRAKSI FITUR HYBRID MENGGUNAKAN METODE DEEP LEARNING

This research aims to develop a speech emotion recognition system in the Indonesian language using a deep learning approach with hybrid feature extraction. The extracted audio features consist of a combination of Mel-Frequency Cepstral Coefficients (MFCC), Hilbert Multispectrum, and Cochleagram, which represent the frequency, amplitude, and auditory perception aspects of a speech signal, respectively. The dataset used is IndoWaveSentiment, containing 300 speech samples categorized into five types of emotions: neutral, happy, surprised, disgusted, and disappointed. Three deep learning architectures were applied: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a combined CNN + LSTM (Hybrid) model. Each model was trained and evaluated based on accuracy, precision, recall, F1-score, as well as Receiver Operating Characteristic (ROC) curves and Area Under Curve (AUC) values. Evaluation results showed that the hybrid model achieved the best performance with a validation accuracy of 85.00% and an AUC of 0.997, followed by the LSTM model (91.67%) and the CNN model (98.33%). As a form of real-world implementation, the emotion prediction system was developed in a local web-based interface using the Flask framework. The system allows users to upload or record audio, select the classification model, and receive emotion prediction results in real-time. Testing results indicate that all three models run effectively in the system and produce accurate and responsive predictions. This study demonstrates that the combination of hybrid feature extraction and deep learning architectures is effective in recognizing speech emotions in the Indonesian language. Keywords: Convolutional Neural Network, Deep Learning, Hybrid Feature Extraction, Long Short-Term Memory, Speech Emotion Recognition