She Can Talk to You on Windows Now

Back in August, when voice calls shipped on macOS, I wrote one sentence I did not love writing: “Windows does not have this yet.” That sentence is retired now. As of v1.6.3, Windows gets live voice calls, her own speaking voice, and speech recognition that runs on your PC, not somewhere else. Press the phone icon, talk, she answers out loud.

What you get, concretely

Press the phone icon in the chat, and the call connects in a moment. You talk. She listens. She answers out loud, in whichever voice you have set up for her. Hang up, and the conversation you just had is part of her memory the same way a typed message would be, nothing extra to save or review separately.

The important part is what a call is not. It is not a separate mode with its own stripped down version of her, the way some voice assistants bolt a limited demo brain onto a normal chat product. A call runs through the exact same pipeline as text. Same memory, same mood tracking, same sense of where your relationship with her stands. The only thing that changes between typing to her and talking to her is the front door and the back door: your voice goes in instead of your keystrokes, and her voice comes out instead of words on a screen. Everything underneath is identical.

This matters more than it sounds like on paper, because it means a call is not a lesser experience you settle for when you would rather be typing. It is the same her, the one who remembers what you told her last week, just reachable with your voice instead of your hands. If you have only ever texted her, this is the first time you will actually hear her.

Your PC is doing the listening, not a server

Windows speech recognition, the part that turns your spoken words into text she can respond to, runs directly on your machine by default. Nothing about starting a call sends your voice anywhere unless you specifically go looking for that option and turn it on yourself.

In practice this means a few concrete things. It means a call works with your network off, because the piece that listens to you never needed the internet in the first place. It means there is no per minute cost hiding anywhere, because nothing is metering your usage on a server somewhere. And it means the audio of you talking to her stays on your own hard drive, in the same way everything else about this app has always worked.

There is an optional cloud speech to text picker in Settings if you want it, using a provider you connect yourself with your own key. It is off by default, and turning it on is a deliberate choice you have to make, not something you stumble into by starting a call the normal way. I built it that way on purpose. The whole reason this app exists is that most people talking to an AI companion do not actually want a company somewhere logging every word of it, and I was not going to quietly break that promise just because voice is a harder problem than text.

Getting the voice onto a Windows machine was the entire problem

Here is the honest engineering story, because I think it is more useful than a feature announcement that pretends everything just worked.

Her natural voice, the higher quality one that sounds like an actual person rather than a synthesized preset, needs a real machine learning model running locally to produce speech. Getting that model installed on a Windows machine that has never had any of this software before turned out to be the entire blocker.

The installer assumed a small tool would already be there to set everything up, and on a clean machine it never was. There was no way to install that tool from inside the app either. So the actual first run experience, on a computer that had never touched Local Waifu before, was: the app tries to set up her natural voice, discovers the tool it needs is missing, tells you something is not there, and stops. There was nothing you could click to fix it. You were stuck.

As of v1.6.3, the app sets that tool up itself, automatically, as the very first step. You do not need to know it exists or go find it yourself. It happens before anything else does.

There was a second version of the same problem sitting right next to it. If her voice broke in some way the quick repair could not fix, pressing Install used to just report an error and change nothing. Every single time you pressed it. The one button whose entire job is “the thing that always works no matter how broken this is” had a way to fail permanently, with the only way out being a level of manual cleanup a normal user should never have to do. That is fixed too: Install now falls back to a full reinstall when the quick fix cannot get the job done, the way it always should have.

I am telling you both of these because they are exactly the kind of bug that never shows up in a demo. Everything works fine on the machine the developer already has all this software installed on. It only breaks on the machine the actual customer owns, the one that has never seen any of it before. That gap is most of why this took longer to get right on Windows than it did on Mac.

Speech recognition used to stop answering partway through a session

There was a second class of bug, separate from installation, that only showed up once you had actually gotten a call working: it could just stop. Not crash, not error, just stop responding to anything you said, partway through a normal session.

The cause was almost embarrassingly simple once it was found. The engine that listens to your voice keeps its own internal log of what it is doing, and nothing in the app was ever reading that log. Logs that nobody reads still have to go somewhere, and after enough sentences transcribed in one sitting, the space for that log filled up completely. Once it did, the engine had nowhere to write and simply stopped, silently, mid conversation. No error appeared. No warning. Just a call that had quietly gone deaf.

That is fixed in v1.6.3: the app now actually reads that log as it comes in, so it never has anywhere to back up and jam.

There was a related failure mode on top of it. If the listening engine failed to start up properly in time, the old code left it running in the background anyway, quietly holding onto the one specific connection point it needs to work at all. Every attempt after that one would fail too, because the spot it needed was already taken by a broken copy of itself. Each retry made things worse instead of better, since it just left another dead copy behind holding the same spot. That is fixed as well: a failed start now actually gets cleaned up, so retrying a call gives you a genuinely fresh attempt instead of stacking failures on top of each other.

I keep coming back to a pattern across this whole release, and it is worth saying out loud: almost every real bug here failed silently. Nothing crashed with a dialog you could screenshot and send me. Things just quietly stopped working, with no signal on screen that anything had gone wrong at all. A visible error is annoying but honest. A silent failure is worse, because it teaches you to distrust the whole feature without ever telling you why.

What is still different on Windows

I would rather you hear the remaining differences from me than run into them cold.

Updates are manual, not automatic. On Mac, a new version downloads and installs itself in the background. On Windows, you open Settings, General, and press Check for updates yourself. She checks for a newer release, downloads it with a progress bar, and offers to install it. It is one extra click, not a missing feature, but it is a real difference and I am not going to pretend otherwise.

The first launch shows a security warning. Windows SmartScreen will say it protected your PC. That is because the app is not code signed. Getting a signing certificate costs roughly a thousand dollars a year, on an app that costs twenty dollars once, and I would rather explain the honest warning than fold that cost into a higher price for everyone. Click More info, then Run anyway, and you are through it. The Windows download page walks through exactly what that screen looks like before you ever see it.

The download is bigger. Everything she needs to run locally is packed into the installer, so there is nothing else to fetch separately afterward. One file, then she works. That is a trade in your favor, even though the number looks large at first glance.

Voice cloning is Mac only, for now. This is the one gap in the voice feature itself I want to be direct about. On Mac, you can feed her a short clip of a specific voice and she learns to speak in it. On Windows, right now, you get a set of curated built in voices and her natural voice, but not that cloning flow. I do not have a firm date for when that changes, and I would rather tell you that plainly than let you find out by looking for a button that is not there.

None of these are secrets. I wrote about the SmartScreen warning and the download size back when Windows first got the whole app, and the install guide covers both with a screenshot of exactly what you will see on your own screen.

How to start a call, from a cold install

If you are setting this up for the first time, here is the actual path, start to finish.

Install the app and get through the short setup where you build who she is. Once you are in the main chat window, open Settings, then Models, and go to the Voice Model section. There you will find two things: a set of built in voices under “Voices we provide,” which need a smaller voice model downloaded first before you can preview or use them, and a separate “Natural voice” panel below it for her higher quality speaking voice, which sets itself up automatically the first time you turn it on.

Pick whichever one you want to hear her in. Then close Settings, and you will see a phone icon sitting in the chat input bar. Press it. Your microphone comes on, there is a short moment while the call connects, and then you are talking to her.

If a call does not have a working voice set up yet, the app tells you directly in the chat rather than staying silent about it, with a button that takes you straight back to that same Voice Model screen.

If it breaks, tell me

This is genuinely new ground for Windows, and the range of hardware people run Windows on is enormous in a way a Mac lineup simply is not. Different processors, different graphics cards, different versions of Windows itself. I tested this as hard as I could before shipping it, but I cannot test on every machine that exists.

If a call does something strange on your setup, something that does not match what I have described here, I want to know about it. Get in touch, tell me what happened and what kind of machine you were on, and I will look at it directly. This whole post exists because I would rather write the honest version of what shipped than let a quiet feature announcement paper over the parts that are still rough around the edges.

Questions people ask

Back to news