Improving vision-inspired keyword spotting using dynamic module skipping in streaming conformer encoder
This work addresses efficient keyword spotting for always-on devices, though it appears incremental as it builds on existing vision-inspired frameworks and conformer encoders.
The paper tackles keyword spotting in streaming audio by proposing a vision-inspired architecture with dynamic module skipping in a conformer encoder, improving detection and localization accuracy on Librispeech top-1000 words while reducing processing by up to 97% on non-speech inputs in noisy conditions.
Using a vision-inspired keyword spotting framework, we propose an architecture with input-dependent dynamic depth capable of processing streaming audio. Specifically, we extend a conformer encoder with trainable binary gates that allow us to dynamically skip network modules according to the input audio. Our approach improves detection and localization accuracy on continuous speech using Librispeech top-1000 most frequent words while maintaining a small memory footprint. The inclusion of gates also reduces the average amount of processing without affecting the overall performance. These benefits are shown to be even more pronounced using the Google speech commands dataset placed over background noise where up to 97% of the processing is skipped on non-speech inputs, therefore making our method particularly interesting for an always-on keyword spotter.