Introducing Y4

Enhanced Radar’s newest aviation understanding model

We’re introducing a new model that expands upon our research to-date in aviation-specific large language models, and our first ground-up, foundation model reimagined for the aviation environment. Y4 was trained de novo on 215 billion aviation tokens (audio + text + geospatial), allowing for robust performance across varying scenarios. Y4 can understand real-world ATC with consistent >98% word level accuracy across 70 U.S. airports, and across facilities such as Local (Tower), Ground, and TRACON.
‍
The model is further designed to accurately distinguish speaker role (e.g. pilot or controller), perform speaker diarization, and to associate transmissions with specific aircraft. All data used to train Y4 was originated and is owned by Enhanced Radar.

Y4 is a new class of proprietary Large Domain Models (LDM). This new model architecture opens the door for continued research that builds solutions specific to aviation’s challenges, which are distinct from the challenges encountered when building AI models for productivity use cases, for instance. Y4 exhibits broad expertise, unlike aviation NLU models to-date which tend to only perform well in constrained scenarios.
‍
This new model is a meaningful step toward systems that work alongside aviation professionals to deliver significant gains in airspace efficiency and safety.
‍
Y4-small, a lightweight preview version of the Y4 model, can be viewed in real-time in the ATC App.