Building ultra low-resource end-to-end ASR models for Indic Languages using the emerging machine-learning paradigm ‘Few-shot Learning’ (FSL).
Project Principal Investigator(s): Dhanya Eledath, PI: Prof. V. Ramasubramanian
In a first-of-its-kind attempt, the project examined and adapted ‘Matching Networks’ (MN), a salient FSL technique belonging to Meta/Metric-Learning for E2E ASR. MN belong to the class of ‘embedding learning’ FSL frameworks, where the network ‘learns to compare’ between the few-shot labeled samples (support set) and test sample (query) in an embedded space.
Proposed Matching Networks – Connectionist Temporal Classification (MN-CTC*) for End-to-end ASR Matching Networks adapted with CTC Loss / Decoding .In a first-of-its-kind attempt, the project examined and adapted ‘Matching Networks’ (MN), a salient FSL technique belonging to Meta/Metric-Learning for E2E ASR. MN belong to the class of ‘embedding learning’ FSL frameworks, where the network ‘learns to compare’ between the few-shot labeled samples (support set) and test sample (query) in an embedded space.Exploits Meta-/Metric-Learning for the first time in ASR.What makes MN / MN-CTC tick?Meta learning (Learning-to learn, Metric learning / Embedding learning,Episodic training,Cross-domain applicability, Test classes different from train classes
