Yes, it does. The Okada Voice Changer DirectML build exists precisely for people without an NVIDIA card, and it lets AMD and Intel graphics do the same real-time conversion work. You download a different build of the same tool, not a lesser imitation of it.
The confusion is understandable, because most talk about the Okada Voice Changer, or VCClient, centres on the CUDA version. CUDA is NVIDIA-only technology, so readers with a Radeon card or Intel graphics often assume the door is closed to them. It is not, and this article explains how the builds divide up and how to choose yours.
Why the Standard Build Leans on NVIDIA CUDA
The original and best-known build of the tool was written around CUDA, NVIDIA’s system for running general computing work on its graphics cards. The AI voice models need exactly that kind of heavy parallel maths, so building on CUDA was the natural first choice.
Because CUDA only exists on NVIDIA hardware, that build simply has nothing to talk to on an AMD or Intel system. The software is not being picky; the technology it depends on is absent from the machine.
Why This Causes So Much Confusion
Download pages and forum posts often mention the tool without naming a build at all. A reader with a Radeon card downloads the famous version, finds it cannot use their graphics card, and concludes the whole project is NVIDIA-only.
In reality the project ships more than one build for a reason. The CUDA build is one door into the same house, and DirectML is another door built for different hardware.
What the DirectML Build Changes
DirectML is Microsoft’s machine-learning layer, built into the DirectX family that Windows already uses for graphics. Instead of depending on one maker’s technology, it can direct AI work to any compatible GPU through a common route.
The DirectML build of VCClient uses that route. The voice models, the conversion, and the way you use the program stay the same, while the behind-the-scenes path to your AMD or Intel graphics changes completely.
Who the DirectML Build Is Really For
It is made for Windows users whose graphics come from AMD or Intel, including a great many laptops and pre-built PCs. If your machine has no NVIDIA card but runs Windows, this is almost certainly your build.
It is also a sensible trial for anyone unsure about their hardware. Because DirectML works across makers, the build is forgiving about exactly which compatible GPU is doing the work.
What You Give Up, and What You Do Not
Choosing DirectML means giving up the small efficiency edge of a technology written for one maker’s cards alone. For most people that is the entire cost, and it is a modest one measured in comfort rather than in features.
You do not give up models, voices, or the live experience itself. Your friends in the call cannot tell which build produced your voice, because the character they hear comes from the model, and the model is identical on every build.
How Performance and Latency Compare, in Honest Terms
NVIDIA’s CUDA route is the mature, well-trodden path, and on comparable hardware it generally squeezes the most out of the models. The DirectML route can carry a little more overhead, which sometimes shows up as slightly higher delay or slightly less headroom with heavy models.
In everyday conversation, though, the difference is usually smaller than people fear. A mid-range Radeon card running the DirectML build can hold a perfectly natural Discord call, and most listeners will never know which build produced the voice they are hearing.
Where the Gap Shows Up Most
The gap matters most at the extremes: very heavy voice models, very weak integrated graphics, or fast competitive games where every fraction of a second counts. That is where honest expectations help, and what latency to expect from each build is worth weighing before you judge the result too harshly.
For relaxed chat, story-driven games, and streaming to a patient audience, both builds land in the same comfortable territory. The deciding factor is more often your choice of model than the name of the build.
Where the Apple Silicon Mac Build Fits
Mac users are not left out of the picture either. The project provides a build for Apple Silicon machines, which uses the graphics and AI hardware inside Apple’s own chips rather than CUDA or DirectML.
That build matters because Mac owners had no route at all in the early days of tools like this. If your computer is a recent Mac with an Apple-designed chip, you ignore both the CUDA and DirectML downloads and take the Mac build instead.
A Separate Build, Not a Fallback
It is better to think of the three builds as siblings than as a ranking. Each one speaks the language of its own platform, and each delivers the same core experience: your voice, converted live, in the character of the model you loaded.
Choosing the Mac build on a Mac is not a compromise, just as choosing DirectML on a Radeon system is not one. The only real mistake is downloading a build your hardware cannot use.
What Stays Exactly the Same Across Builds
Whichever build you download, the program you open is recognisably the same tool. The layout, the way models are loaded, and the daily routine of choosing a voice and pressing start do not change between the CUDA, DirectML, and Mac versions.
The sound you produce comes from the same place too: the voice model you loaded. Two friends using different builds with the same model will sound far more alike than two friends sharing a build with different models.
Your Settings and Habits Carry Over
Everything you learn on one build remains useful on another. Microphone technique, input selection in Discord, and judging pitch by ear are identical skills whether the maths runs through CUDA or DirectML.
That also means advice from other users mostly applies to you, even when their hardware differs. A tip about routing, monitoring, or model choice rarely depends on the build, so the wider community’s experience is yours to borrow.
How to Pick the Correct Build for Your Own Machine
Start with one question: who made your graphics hardware? If the answer is NVIDIA, the CUDA build is your default. If the answer is AMD or Intel on a Windows PC, take the DirectML build, and if you are on an Apple Silicon Mac, take the Mac build.
If you do not know what is inside your machine, your system settings will tell you. The display or about section of any modern PC lists the graphics hardware by name, and that single fact settles the download choice completely.
A Simple Way to Decide in Under a Minute
Check the name of your graphics card, match it to the build for that maker, and download that one first. You can always try a different build later, but starting with the build written for your hardware avoids nearly every early disappointment.
One last caution: download builds from the project’s own pages rather than random re-uploads. Voice changers are popular targets for tampered copies, and a clean download from the source protects both your machine and your accounts.
If Your Laptop Has Two Graphics Chips
Many laptops carry both integrated graphics and a separate gaming chip, and the maker’s name on the stronger chip is the one that decides your build. If that chip is NVIDIA, start with CUDA even though Intel graphics are also present; if the machine is AMD and Intel only, DirectML remains your route.
Frequently Asked Questions
Will my RVC voice models work in the DirectML build?
Yes. The models are the same files across builds, because the build only changes how the conversion work reaches your graphics hardware. A model you used on the CUDA build can be loaded in the DirectML build in the same way.
Is the DirectML build noticeably worse than the CUDA build?
Usually not in ordinary conversation. It can carry slightly more overhead, which may add a little delay with heavy models, but for typical Discord chat and casual gaming most people find the experience comfortably similar.
Can Intel integrated graphics run it at all?
In many cases, yes, through the DirectML build, though integrated graphics have the least headroom of any option. Light voice models and relaxed expectations make the difference between a frustrating trial and a pleasant surprise.
Do I need to remove anything before switching builds?
No deep cleanup is normally required, because the builds are separate downloads of the same tool. Keep your model files somewhere safe, download the build that matches your hardware, and point it at the models you already have.
Conclusion
Not owning an NVIDIA card does not lock you out of this voice changer. The CUDA build serves NVIDIA hardware, the DirectML build opens the same live conversion to AMD and Intel graphics on Windows, and the Apple Silicon build covers modern Macs.
Match the build to the name on your graphics hardware, choose a model your machine can handle, and the voice your friends hear will be every bit as entertaining as the one from the famous CUDA version. The build is a doorway, and there is one built for you.