tinyML Asia 2021
Partner Session – Qualcomm
Always-on audio/speech network architectures, personalization, and framework
Kyuwoong HWANG, Senior Director, Qualcomm Research, Korea
We introduce efficient network architectures for audio/speech always-on applications such as keyword spotting and acoustic scene classification that leverages domain characteristics to achieve good performance with minimal memory and computation. Personalized pruning at a device with only a forward pass reduces a network when a specific person uses it, which is typical in many cases. We briefly introduce our deep machine learning framework, too.