Deep Learning-Based End-to-End Speaker Identification Using Time-Frequency Representation of Speech Signal

Speech-based speaker identification system is one of the alternatives to the conventional biometric contact-based identification systems. Recent works demonstrate the growing interest among researchers in this field and highlight the practical usability of speech for speaker identification across various applications. In this work, we try to address the limitations in the existing state-of-the-art approaches and highlight the usability of convolutional neural networks for speaker identification systems. The present work examines the usage of spectrogram as an input to these spatial networks and its robustness in the presence of noise. For faster training (computation) and to reduce the memory requirement (storage), SpectroNet model for speech-based speaker identification is introduced in this work. Evaluation of the proposed system has been done using Voxceleb1 and Part1 of the RSR 2015 databases. Experimental results show a relative improvement of similar to 16% (accuracy-96.21%) with spectrogram and similar to 10% (accuracy-98.92%) with log Mel spectrogram in identifying the speaker compared to the existing models. When cochleagram was used, it results in an identification accuracy of 99.26%. Analyzing the result obtained shows the applicability of the proposed approach in situations where (i) minimal speech data are available for speaker identification; (ii) speech data are noisy in nature.