Online Continual Learning in Keyword Spotting for Low-Resource Devices via Pooling High-Order Temporal Statistics

TL;DR

This paper introduces a novel online continual learning method for keyword spotting on resource-limited devices, using high-order temporal statistics to improve incremental recognition without retraining large models.

Contribution

It proposes Temporal Aware Pooling (TAP) and TAP-SLDA, a new approach that leverages high-order moments for effective incremental learning in embedded KWS systems.

Findings

TAP-SLDA outperforms existing methods on multiple benchmarks.

Achieves an average of 11.3% relative gain on GSC dataset.

Effective in resource-constrained embedded environments.

Abstract

Keyword Spotting (KWS) models on embedded devices should adapt fast to new user-defined words without forgetting previous ones. Embedded devices have limited storage and computational resources, thus, they cannot save samples or update large models. We consider the setup of embedded online continual learning (EOCL), where KWS models with frozen backbone are trained to incrementally recognize new words from a non-repeated stream of samples, seen one at a time. To this end, we propose Temporal Aware Pooling (TAP) which constructs an enriched feature space computing high-order moments of speech features extracted by a pre-trained backbone. Our method, TAP-SLDA, updates a Gaussian model for each class on the enriched feature space to effectively use audio representations. In experimental analyses, TAP-SLDA outperforms competitors on several setups, backbones, and baselines, bringing a relative average gain of 11.3% on the GSC dataset.