BlackRook Posted June 3 Posted June 3 Is it still true that the MacOS version is the best version? Or have those updates made their way into the other OS versions, as well?
intruder11 Posted June 29 Author Posted June 29 (edited) On 6/3/2025 at 7:40 AM, BlackRook said: Is it still true that the MacOS version is the best version? Or have those updates made their way into the other OS versions, as well? The macOS version of SkyEye uses features of macOS (Metal, Speech Synthesis Manager) and the Apple Silicon CPUs (M1/M2/M3/M4) that are not available on the other operating systems. The parts of the code that make SkyEye great on macOS will therefore never be available on Windows and Linux. In theory, I could achieve similar levels of speech recognition performance for people who own certain Nvidia graphics cards (and on Linux, use the proprietary driver from Nvidia instead of the open Nouveau driver) by using CUDA. However, I don't own any Nvidia hardware to test on. That might change late this year/early next year if I build another rig, but for now all my Windows/Linux machines use AMD GPUs. The Speech Synthesis Manager has no direct equivalent on any other OS - technology of that quality simply does not exist outside of macOS. I've tried literally every alternative available. Even the hosted speech models from companies like OpenAI don't compare. (These Apple CPUs are so good. I hadn't bought a laptop for myself in over ten years - after work provisioned me an M4 Macbook, I went out and bought two for myself, plus a Mac Mini that's currently running SkyEye 24/7 for Flashpoint Levant. They're amazing hardware, the best non-gaming computers you can buy for the money IMO.) Edited June 29 by intruder11
Caput_58 Posted September 16 Posted September 16 (edited) Just a quick note to say how impressed I was with the relative ease of setting up Skyeye on a home machine. It was less than 30 mins work, mostly because I couldn't remember how to change my server's Tacview settings. I'm still shocked that it hasn't been adopted by more mp servers. When it's processing a transmission, I notice the CPU usage spikes up to 100%. Does that mean Skyeye is spreading the load effectively over all cores? Assuming I stick to local processing on a PC, would it benefit from a threadripper or other CPU that's big on core count? Does anyone else have other ideas on optimizing response time other than using the quicker language model? Thanks again for all the work. It's another example of the strong community support that brings life to DCS. Oh, I just remembered my other question. Is Skyeye impaired at all by that radio effects simulation of SRS? Surely decreased sound quality of a simulated radio transmission is making Skyeye work harder to understand what's said. Edited September 16 by Caput_58 addendum
intruder11 Posted 8 hours ago Author Posted 8 hours ago (edited) I've released v1.5.0. Highlights include: MiG-29 Full Fidelity support, SRS recording support, and some minor improvements to speech recognition quality. https://github.com/dharmab/skyeye/releases/tag/v1.5.0 On a personal note, development is slower now because of other stuff in my life being quite busy (my professional work, home improvement projects, non-digital hobbies, and family time). I hope to resume major development in the future but for now expect mostly maintenance rather than big new features. Quote When it's processing a transmission, I notice the CPU usage spikes up to 100%. Does that mean Skyeye is spreading the load effectively over all cores? Assuming I stick to local processing on a PC, would it benefit from a threadripper or other CPU that's big on core count? Yes, when you run local speech recognition on Windows or Linux, all available CPU cores are used, and newer, faster CPUs perform better. See the hardware section of the admin guide: https://github.com/dharmab/skyeye/blob/main/docs/ADMIN.md#hardware Quote Does anyone else have other ideas on optimizing response time other than using the quicker language model? Some people have made forks of SkyEye that can use GPU acceleration on Windows and Linux, using CUDA or Vulkan. I'm interested in adding support for this in the standard build, but have two obstacles: 1. I don't have any nvidia GPUs to test CUDA 2. Limitations on my time mentioned above I would strongly consider using the OpenAI API, if possible; it's _really_ cheap and works great. Running on local CPU should be considered a last option if a Mac or the OpenAI API is not possible. Quote Is Skyeye impaired at all by that radio effects simulation of SRS? No, SkyEye bypasses all SRS effects. SRS radio effects are client-side; you can customize them or turn them off in the official client. SkyEye uses custom code instead of an official SRS client, and simply does not implement any audio effects, instead directly converting the SRS network packets' data into the format used by the AI models. Edited 8 hours ago by intruder11 1
Recommended Posts