Speaker Recognition on Low Power Device Using Fully Convolutional QuartzNet

The need for a small and lightweight algorithm used for speaker recognition that can run on low-power devices is on the rise. This is mainly caused by security and privacy concerns of users with the use of their personal and biometric data. The speaker recognition task is mainly used as a biometric authentication, so an accurate model is also needed. The previous method uses feature engineering to extract features from raw audio files with heavy reliance on the training data and a dissimilarity between the training data and real-world implementation causes a significant decrease in its accuracy. We propose a Fully Convolutional QuartzNet as a deep learning approach to this problem. We achieved 84.6% accuracy when testing on a small subset DR-VCTK dataset with 30 classes and 56.40% accuracy on a small subset of the VoxCeleb dataset with fewer files for each of the 125 classes. The proposed model was also tested for binary speaker recognition, achieving 5.07% EER. We also achieve a small parameter count of only 33K parameters without sacrificing significant performance, and the proposed method can achieve its highest accuracy with only 53K parameters.