FSIR: Few-Shot Speaker Identification using Reptile Algorithm
Banala Saritha, Anish Monsley Kirupakaran, Rabul Hussain Laskar, Madhuchhanda Choudhury
Abstract
Speaker identification (SI) is the evolving biometrics and holds great potential for real-world applications. SI is challenging task with limited data in forensics. Few-shot learning inspires from human learning. In this work, SI tasks are investigated using optimization based meta-learning. A novel architecture for SI system is proposed and is based on a convolutional recurrent neural network. The accuracy is improved by 3% using the proposed framework, outperforming statistical methods.
Topics & Concepts
Computer scienceIdentification (biology)Shot (pellet)Speaker identificationSpeech recognitionArtificial intelligenceAlgorithmSpeaker recognitionPattern recognition (psychology)BiologyChemistryBotanyOrganic chemistrySpeech and Audio ProcessingSpeech Recognition and Synthesis