This project implements a Speech Emotion Recognition system that classifies emotions from speech samples using deep learning models. The project uses four datasets: CREMAD, RAVDESS, SAVEE, and TESS. The model achieves an accuracy of 96% by combining CNN, LSTM, and CLSTM architectures, along with data augmentation techniques and feature extraction methods.
The following datasets are used for training and testing:
- CREMAD: A dataset of emotional speech for automatic detection.
- RAVDESS: The Ryerson Audio-Visual Database of Emotional Speech and Song.
- SAVEE: The Surrey Audio-Visual Expressed Emotion (SAVEE) dataset.
- TESS: Toronto emotional speech set, consisting of speech data with various emotions.
- MFCC (Mel-frequency cepstral coefficients)
- Zero Crossing Rate
- RMS (Root Mean Square)
- Time Shifting
- Pitch Shifting
- Time Stretching
- Noise Addition
- CNN (Convolutional Neural Networks)
- LSTM (Long Short-Term Memory)
- CLSTM (Convolutional LSTM)
These models were trained on the combined dataset and evaluated for emotion classification.
The model achieved 96% accuracy after training.
-
Import into Kaggle:
- Go to the Kaggle and import this notebook into your Kaggle environment.
-
Add Datasets:
- Upload the datasets (CREMAD, RAVDESS, SAVEE, TESS) to the Kaggle notebook by navigating to the “Datasets” tab and adding them to your notebook.
-
Run the Notebook:
- Run the notebook step-by-step to preprocess the data, perform feature extraction, apply data augmentation, and train the models.
- Ensure that the necessary libraries are installed in the Kaggle environment.
-
Training:
- Once the datasets are loaded and prepared, you can run the training code to train the CNN, LSTM, or CLSTM models and evaluate the performance.
The model achieves an accuracy of 96% on the evaluation dataset, with the following confusion matrix and performance metrics:
