Small Footprint Convolutional Recurrent Networks for Streaming Wakeword Detection

TL;DR

This paper introduces small footprint convolutional recurrent neural networks with attention for streaming wakeword detection, achieving significant reductions in false accepts and model size while maintaining accuracy.

Contribution

It presents a novel application of CRNNs with attention to wakeword detection, demonstrating improved performance and efficiency over traditional CNN and DNN models.

Findings

25% reduction in false accepts with 10% fewer parameters at 250k budget

Up to 32% improvement at 50k parameter budget

Effective inference solutions for streaming audio with CRNNs

Abstract

In this work, we propose small footprint Convolutional Recurrent Neural Network models applied to the problem of wakeword detection and augment them with scaled dot product attention. We find that false accepts compared to Convolutional Neural Network models in a 250k parameter budget can be reduced by 25% with a 10% reduction in parameter size by using CRNNs, and we can get up to 32% improvement at a 50k parameter budget with 75% reduction in parameter size compared to word-level Dense Neural Network models. We discuss solutions to the challenging problem of performing inference on streaming audio with CRNNs, as well as differences in start-end index errors and latency in comparison to CNN, DNN, and DNN-HMM models.