Emotion Recognition from Speech – an LSTM approach with the Tess Dataset

The research provides the analysis of speech emotions using a dataset comprising audio recordings, leveraging the TESS dataset and Long Short-Term Memory (LSTM) networks. Overcoming resource-intensive challenges and limited emotion-labeled data, the work enhances the efficiency and effectiveness of emotion recognition systems. The TESS dataset is selected for its diverse emotional speech samples, enabling comprehensive model training and evaluation. This methodology employs Mel-frequency cepstral coefficients (MFCCs) for acoustic feature extraction, with LSTM networks as the core architecture. LSTM’s ability to capture temporal dependencies in speech data enhances emotion recognition accuracy. Model training includes hyperparameter optimization for consistent performance across various emotions. This approach, while offering a viable method for speech emotion recognition, acknowledges that it may not surpass the performance of certain existing methods. Nonetheless, it addresses resource constraints and demonstrates improvements in accuracy compared to previous approaches. This approach effectively addresses resource constraints and exhibits improved accuracy compared to prior approaches. The evaluation process ensures the reliability and validity of the model’s performance with real-world scenarios. The model achieved a notable accuracy of 91.45% in recognizing the emotions. By contrasting the anticipated labels in the test set with the genuine labels, this accuracy is determined. Model training includes hyperparameter optimization for consistent performance across various emotions.