Speaker Recognition Using Noise Robust Features and LSTM-RNN

A tremendous growth has been observed in terms of active research in the field of speaker recognition. This has been mainly due to the increasing need of zero-touch interfaces in devices and mobile biometric authentication systems. This paper discusses implementation of text-independent speaker verification system using long short-term memory (LSTM)-based neural network for speaker modeling by using various approaches for the front-end feature extraction including Mel Frequency Spectral Coefficients (MFSC), Mel Frequency Cepstral Coefficients (MFCC), Gammatone Filter Spectra (GTF), and Gammatone Filter Cepstral Coefficients (GFCC). Additionally, to determine the best-suited speaker verification system for given noisy conditions of environment, all the combinational systems are tested under induced noisy conditions with white noise at −20 and −40 dB, as well as under clean environmental condition. The results show that the MFSC-based LSTM-RNN combination tends to perform better than all the other combinations regardless of the noise added in the dataset.