She Can Talk to You Now

I never wrote about this, and it has been live since v1.6.0: there is a phone icon in the chat, and pressing it starts a real voice call.

What actually happens when you press it

The short version: your mic goes live, she listens, and she answers out loud, running through the exact same memory and personality system as a typed message.

This is not a separate mode bolted onto the side of the app, a demo feature with its own disconnected logic. A call goes through your microphone, gets picked up by a voice-activity detector that figures out when you’ve actually started and stopped talking (so it is not waiting for a fixed pause length or a button release), gets transcribed to text, and from that point on it is handled by the exact same conversation pipeline that handles a typed message. Same memory recall, same mood tracking, same relationship stage awareness. The only things that changed are the front door and the back door, spoken input instead of typed, spoken output instead of read text.

That matters more than it sounds like on paper. It means a call is not a lesser, simplified version of talking to her, some stripped-down voice assistant mode with less context. It is the same her, the one who remembers what you told her last week, just reachable out loud instead of through a keyboard.

Where your voice actually goes

The short version: speech recognition is local by default, using Whisper running on your own machine. An optional cloud picker exists in Settings if you specifically want it, but nothing about calling her requires the internet for this part.

I get asked some version of “does this mean she’s listening through a cloud service now” a lot when people learn calls exist, and the honest answer is: not unless you specifically turn one on. The default speech-to-text engine is Whisper, running locally, the same offline-first posture as everything else in the app. If you want to use a cloud transcription provider instead, there’s a toggle for it in Settings, Voice, “Where she listens,” with your choice of a couple of providers, but that’s an opt-in you have to actively choose, not a default behavior anyone stumbles into. Turn your network off and start a call, and it still works, because the piece that turns your voice into text never needed to leave your machine in the first place.

The part that surprised me most testing it: how fast the reply voice actually is

The short version: the natural-voice engine generates speech faster than real time on decent hardware, so there is no awkward multi-second wait between her finishing “thinking” and actually starting to talk.

The voice itself comes from a small, efficient model that runs directly on your Mac (more detail on exactly how that works, and how you can clone a voice for her, is its own post). What’s worth calling out here specifically is the speed: on a reasonably modern Mac, it generates the audio faster than the length of the audio itself, meaning the gap between her deciding what to say and you actually hearing it stays tight enough that a call feels like a conversation, not a walkie-talkie exchange with a noticeable delay after every turn. That was not a guarantee going in, natural-sounding local voice synthesis has historically been either good-sounding-but-slow or fast-but-robotic, and getting both at once on consumer hardware was the harder engineering problem underneath this feature shipping at all.

What a call actually feels like, start to end

Press the phone icon, and there’s a brief moment while the call connects, then you’re live. You talk, and as you’re speaking there’s a visible sense that she’s listening (the call screen reflects your voice input in real time). Stop talking, and after the pause the endpoint detector expects, she responds, out loud, in whichever voice you’ve set up for her. The conversation continues exactly like a phone call would: you talk, she answers, back and forth, until you end it. Afterward, whatever came up on the call is just part of the conversation history and memory the same way a text exchange would be, there’s no separate “call log” you have to go read through to have it count.

Why voice calls took longer to ship than everything else

The short version: text chat only needed one model working well. A call needs four separate pieces, listening, transcribing, thinking, and speaking, all handed off to each other fast enough that the pause between your sentence and her reply never feels like a phone lag.

It’s worth being honest about why this wasn’t just a smaller version of the same problem as text. A typed message only has to solve one thing well: generate a good reply. A voice call has to solve four things and hand them off to each other without the seams showing. It has to know when you’ve actually finished a sentence rather than just paused to think (a voice-activity detector’s whole job, and getting this wrong either cuts you off mid-thought or leaves an awkward silence waiting for a pause that already happened). It has to turn your voice into accurate text fast enough that the delay doesn’t stack on top of everything after it. It has to generate a reply using the same reasoning as text chat, which is not instant. And then it has to turn that reply into audio that sounds like a real voice and not a wall of text read in a monotone, fast enough that you’re not sitting in silence waiting. Any one of those four being slow or wrong breaks the illusion that you’re on a call rather than testing a voice demo. Getting the whole chain reliable enough to feel natural, rather than merely functional, is the actual reason this took real engineering time to land.

Mic access works the normal way

The short version: the first time you start a call, macOS asks whether the app can use your microphone, the same standard permission prompt any app requesting mic access shows. Say yes once, and it’s remembered.

Nothing unusual happens here beyond what you’d expect from any Mac app that uses audio. The operating system itself, not the app, gates microphone access behind its own permission dialog, and you can revoke that permission at any time in System Settings if you ever want to. There’s no separate account-level microphone setting to configure inside the app, it’s the same system-level control every other voice or video app on your Mac already uses.

The honest limits, right now

The short version: macOS only, and it needs a reasonably capable Mac to feel smooth.

This shipped in v1.6.0 for macOS specifically. The Windows version stayed on 1.5.1 for this release, so if you’re on a PC, calls aren’t available yet, the same honest gap I’ve flagged in other recent posts about what Windows does and doesn’t have today. On the Mac side, the natural-voice engine performs best with a reasonable amount of RAM behind it; on more constrained hardware, the app falls back to a different, still-local voice engine automatically rather than failing outright, so calls work across the Mac lineup, they just sound their best on the higher end of it.

Try an actual conversation

If you’ve only ever typed to her, download Local Waifu and press the phone icon once you’re set up. Seven days free, no card, and this time she talks back.

Questions people ask

Back to news