Bringing Frictionless Speaker Verification to the Edge: An ASV Solution for Indian Languages

For widespread adoption in business use cases, an Automatic Speaker Verification (ASV) system must deliver both high accuracy and real-time performance. To enhance ASV accuracy, we fine-tuned a state-of-the-art ResNet34-based ASV model for Hindi and Kannada, improving feature discrimination and achieving a 1.62 % reduction in Equal Error Rate (EER) for Hindi and a 1.12% reduction for Kannada language. We employ a fine-tuning process to enforce higher intraclass similarity and greater inter-class discrimination using Additive Angular Margin (ArcFace) loss and Triplet loss. For real-time deployment, we implemented these fine-tuned models on edge devices (Android) using the Sherpa-ONNX framework. By eliminating network dependency, our Android deployment ensures real-time verification while also reducing the cost associated with server-based solutions.