Ethnicity Prediction From Audio Samples for Different Languages Using Convolutional Neural

“Ethnicity Prediction from Voice Samples” deals with prediction of a person's ethnicity from his/her voice samples. Ethnicity Prediction has gained more and more importance and attention in fields like security, recommendations systems, learning apps etc. The proposed work uses a customized Convolutional Neural Network (CNN) based model to predict the ethnicity of the speaker based on combination of varied audio features like Mel-Frequency Cepstral Coefficients (MFCC), pitch, energy, and zero crossing rate. The different voice samples are obtained from the popular Speech Accent Archive Dataset. Augmentation of data by performing time and pitch shifting proved to give a significant boost in the accuracy as it increased the number of voice recordings. The work also compares the performance of the proposed model with other state-of-art models like LSTM. The F-score obtained was 98% for the class Russian, Polish and 90% for the class French, Italian, Spanish i.e. Romance Languages and 80% for mixed language category that comprises of Indian, Italian, German and Polish.