SkyEye is a Ground Controlled Intercept (GCI) bot for the flight simulator Digital Combat Simulator (DCS). A GCI bot allows players to request information about the airspace in English using either voice commands or text entry, and to receive answers via verbal speech and text messages
SkyEye uses Speech-To-Text and Text-To-Speech technology which runs locally on the same computer as SkyEye. No cloud APIs are required. It works with any DCS mission, singleplayer or multiplayer. No special scripting or mission editor setup is required. You can run it for less than a nickel per hour on a cloud server, or run it on a PC in your home.
SkyEye is under active development. All of the radio calls I planned to support have been implemented - but there is still lots of work to do on performance, quality, accessibility, and additional features. To see what I'm working on, check out the milestones!
- Players: See the user guide (work in progress) for instructions on using the bot.
- Server admins: See the admin guide (work in progress) for a technical guide on deploying the bot.
- Developers: See the contributing guide for instructions on building, running and modifying the bot.
- Please also see the privacy statement to understand how SkyEye uses your voice and gameplay data to function.
Skyeye would not be possible without these people and projects, for whom I am deeply appreciative:
- DCS-SRS by @ciribob. Ciribob also patiently answered many of my questions on SRS internals and provided helpful debugging tips whenever I ran into a block in the SRS integration.
- Tacview - specifically, ACMI real time telemetry - provides the data feed from DCS World.
- @rurounijones's OverlordBot was a useful reference against SkyEye during early development, and Jones himself was also patient with my questions on Discord.
- @ggerganov's whisper.cpp models provides text-to-speech.
- @rodaine's numwords module is invaluable for parsing numeric quantities from voice input.
- Piper by the Rhasspy voice assistant project is used for speech-to-text.
- The Jenny dataset by Dioco provides the feminine voice for SkyEye.
- @popey's dataset provides the masculine voice for SkyEye.
- @amitybell's embedded Piper module makes distribution and implementation of Piper a breeze. @nabbl improved this module by adding support for macOS and variable speeds.
- The Opus codec and the
hraban/opus
module provides audio compression for the SRS protocol. - @hbollon's go-edlib module provides algorithms to help SkyEye understand when it slightly mishears/the user slightly misspeaks a callsign or command over the radio.
- @lithammer's shortuuid module provides a GUID implementation compatible with the SRS protocols.
- @zaf's resample module helps with audio format conversion between Piper and SRS.
- @martinlindhe's unit module provides easy angular, length, speed and frequency unit conversion.
- @paulmach's orb module provides a simple, flexible GIS library for analyzing the geometric relationships between aircraft.
- @proway's go-igrf module implements the International Geomagnetic Reference Field used to correct for magnetic declination.
- Cobra is used for the CLI frontend, including configuration, help and examples.
- MSYS2 provides a Windows build environment.
- Oto was helpful for debugging audio format conversion problems.
- zerolog is helpful for general logging and printf debugging.
- testify is used in unit tests.
- Multiple DCS communities provide invaluable feedback and morale-booster energy:
- Team Lima Kilo and the Flashpoint Levant community
- The Hoggit Discord server
- Digital Controllers
- 1VSC
- CVW8
- @Frosty-nee
- The Ace Combat series by PROJECT ACES/Bandai Namco and Project Wingman by Sector D2 are massive influences on my interest in GCI/AWACS, and aviation in general. This project would not exist without the impact of Ace Combat 04: Shattered Skies.
- And of course, DCS World is produced by Eagle Dynamics.
This project is currently available in Limited Availability. Anyone can download the software and try it out, but it may contain bugs or have performance or quality issues. I am currently only providing support for a limited number of personal friends.
A General Availability release is expected during winter 2024-2025. At that point, I expect the software to be stable with few or no issues, and support will be provided to the general audience.
You can check current progress here!
CPU: SkyEye's speech recognition is extremely sensitive to CPU latency. It does not run well when sharing a CPU with other intensive software.
- Avoid running SkyEye on the same physical machine as another intensive app like DCS or TacView client. Ideally, run it on a separate computer.
- If you're running SkyEye on a cloud provider, ensure your virtual machine has dedicated CPU cores instead of shared CPU cores.
- SkyEye is heavily multi-threaded and benefits from multi-core performance.
Memory: SkyEye uses about 2.5-3.0GB of RAM when using the ggml-small.en.bin
model.
Disk: SkyEye requires around 1-2GB of disk space depending on the selected Whisper model.
Some examples of the performance you can expect:
- My personal rig: AMD 5900X, 64GB DDR4 RAM. Speech recognition takes 1.5-3.0 seconds.
- Hetzner CCX23: AMD EPYC Milan (4 dedicated cores), 16GB RAM. Speech recognition takes around 5-6 seconds.
- Hetzner CCX13: AMD EPYC Milan (2 dedicated cores), 8GB RAM. Speech recognition takes around 13-16 seconds.
Since the software runs 100% locally, the speech recognition model is a local file. Server operators can provide a trained model as an alternative to the off-the-shelf model. See this blog post for an example.
I don't plan to provide a mechanism for players to submit their voice recordings to the main repository due to data privacy concerns.
No. Excluding this feature was an explicit choice in order to avoid the complexity demon.
If this is a critical feature for you, consider using MOOSE's AWACS module instead. It supports Line-Of-Sight and datalink simulation, at the tradeoff of requiring some special setup in the Mission Editor.
OverlordBot also optionally supports this feature, although less than 1% of users used it.
Hopefully in the future Eagle Dynamics will add support for external GCI bots. If anyone at ED is reading this, access to any relevant preview builds would be really helpful!
This deserves a longer answer, for now see this issue
TL;DR most of the controller logic is simple geometry that completes in about a millisecond. An LLM is several orders of magnitude slower, less accurate and a more difficult user experience.
We use AI for the "squishy" problems - understanding human speech, and synthesizing human-like speech. We use traditional code for the algorithmic problems.
This deserves a longer answer, for now see this issue
TL;DR I have no plans to attempt an ATC bot.
October 12th. At some point I'll put an Ace Combat 04 easter egg in there.