Computing with Hypervectors for Efficient Speaker Identification

TL;DR

This paper presents a simple, fast, and resource-efficient hypervector-based method for speaker identification that achieves competitive accuracy with minimal parameters and computational resources.

Contribution

The paper introduces a novel hypervector computing approach for speaker ID that significantly reduces complexity while maintaining high accuracy.

Findings

Achieves 31% Top-1 accuracy with minimal parameters

Improves to 48% Top-1 accuracy with additional training

Classifies 1 second of speech in 5.7 ms on CPU

Abstract

We introduce a method to identify speakers by computing with high-dimensional random vectors. Its strengths are simplicity and speed. With only 1.02k active parameters and a 128-minute pass through the training data we achieve Top-1 and Top-5 scores of 31% and 52% on the VoxCeleb1 dataset of 1,251 speakers. This is in contrast to CNN models requiring several million parameters and orders of magnitude higher computational complexity for only a 2$\times$ gain in discriminative power as measured in mutual information. An additional 92 seconds of training with Generalized Learning Vector Quantization (GLVQ) raises the scores to 48% and 67%. A trained classifier classifies 1 second of speech in 5.7 ms. All processing was done on standard CPU-based machines.