First of all, you need to install keras and tensorflow. In additon, you need to install other dependencies: sklearn and pickle. You need to extract the features and save them as pickle file. I used spectrogram as feature here but it can be replaced with other feature types. 1. Store your audio filenames in file: training_labels.lst In this file, the first column is filename and second column is the label. 2. The validation trials are in file: VoxCeleb-1_validation_trials.txt 3. To train a cnn model in keras, run: "python CNN.py --train 1 --predict 0" a. This will load data from the path you provide in the CNN.py script. b. It will train the model. c. It will save the trained model in the same directory. 4. To extract speaker embeddings using the trained model and score trials, run: "python CNN.py --train 0 --predict 1" The scores will be saved in text file: scores_VoxCeleb-1 5. Finally to compute Equal Error Rate, type: "python EER.py ./scores/scores_VoxCeleb-1"