This presentation deals with the integration of Julius Speech Recognition Engine.

The aim of this Proof of Concept is to have a connectionless speech engine, working on an embedded device, integrated as a binding of the AGL Application Framework. The recognition uses Deep Neural Network realtime decoding, and for safer results and performances purpose, uses a grammar.

Julius does not support wakewords out of the box, some hacking has been done to enable it in an efficient way. Tests have been done on Renesas' H3, and UPSquare boards

Download slides [click here]