ReptoNet: A 3D Log Mel Spectrogram-Based Few-Shot Speaker Identification with Reptile Algorithm

A speech-oriented speaker identification system offers an alternative approach to conventional biometric identification systems that rely on physical contact. Recent advances in deep learning have yielded impressive results with large amounts of data, but they are ineffective in forensic and law enforcement applications due to insufficient data. Consequently, speaker identification tasks may not always be feasible with limited data. Recent developments in meta-learning have opened the door to numerous few-shot learning applications. Nevertheless, significant challenges remain, including limited data, noisy and multivariable input conditions, overfitting, and the need for generalization to unseen speakers. This paper introduces a meta-learning-based speaker identification system utilizing a new framework named ReptoNet, which uses three-dimensional (3D) log Mel spectrogram inputs to minimize overfitting and improve generalization. ReptoNet utilizes a reptile algorithm to perform speaker identification tasks and uses the Keras framework for the implementation. The system is evaluated against state-of-the-art techniques on three diverse speech databases: VCTK, VoxCeleb1, and IIT Guwahati multivariability (IITG-MV). It outperforms existing methods on VCTK, improving log Mel and 3D log Mel inputs by 7% and 6%, respectively, and VoxCeleb1 by 2%.