Fast binary features for speaker recognition in embedded systems

Abstract

Due to the growth in popularity of low power embedded devices with integrated biometric data and algorithms for user verification; speaker recognition problem with limited resources becomes more and more of an interest. Limited memory and processing power of embedded devices requires simple yet robust speaker recognition algorithm. In this work the influence of histograms of fast binary features from raw audio signal on speaker recognition is investigated. Librispeech dataset with audio records is being analyzed. Histograms of binary features are calculated on raw audio data and set of histograms' parameters and distance functions is applied to minimize recognition errors by calculating false acceptance, false rejection error rates and ranking. Preliminary results show that this approach can provide time invariant features and low calculation costs with equal error rate up to 12.71% on Librispeech test-clean corpus.