Efficient Low-Complexity Speaker Identification based on a SpeakerGAN Approach

This study examines the SpeakerGAN-based model for speaker identification with the aim to lower memory requirements enabling its training on low-power computers. By employing an encoder/decoder structure within SpeakerGAN, the proposed approach augments data by reconstructing samples. This augmentation results in increased accuracy, particularly when dealing with smaller databases. The proposed Speaker-GAN modification also speeds up convergence, thus enabling training on low-performance computers. The experimental studies on the LibriSpeech dataset prove the ability of GANs to enhance system accuracy with constrained data.