Multi-Lingual Digital Assistance on Edge Devices - A Hybrid Model Approach

A conference talk from the tinyML Summit 2023 explores the development of hybrid multilingual digital assistance systems for edge devices. Learn how Cisco Systems approaches the challenge of reducing latency in speech recognition by combining cloud and edge computing capabilities. Discover the innovative architecture of a hybrid model that includes a short-phrase "local command-recognition" system for edge devices and a comprehensive natural language command-recognition system in the cloud. Explore the technical aspects of model training, including the implementation of a common embedding network for multiple languages and language-specific decision networks. Understand the process of command selection for edge device processing and the resulting improvements in user experience through reduced latency. Follow the complete development journey from concept to implementation, covering topics such as local commands architecture, modeling user intents, and future directions in edge-based digital assistance.