On-device voice, and phone calls that actually ring.

I drafted most of this post out loud, pacing around the room with the keyboard untouched. Speaking is the fastest input most of us own, and typing has held on as the default mostly out of habit. The reason voice never stuck for real work is that it always shipped with two strings attached: your audio went to somebody else's server, and the thing listening could talk back but could not act. OMEGA's voice stack exists to cut both strings.

So this is a deep dive on two halves. The first is on-device speech: recognition and synthesis that run entirely on Apple Silicon, offline if need be, with no audio ever leaving your Mac. The second is the part I have not seen any other desktop assistant ship: OMEGA can pick up a real phone line. Reservations, follow-up calls, appointment scheduling. The calls you keep postponing. I will also be specific about where this fails and when you should still dial yourself, because a voice feature that oversells itself is worse than no voice feature at all.

Speech that never leaves the machine

When you talk to OMEGA, a speech model running locally on your Mac turns the audio into text. When OMEGA answers out loud, the synthesis is local too. There is no upload step and no round trip to a transcription API. The audio exists on your machine and nowhere else, and the text it becomes follows the same rules as anything you type.

Two things fall out of that design. The first is that it works offline. On a flight with no Wi-Fi you can still dictate notes into OMEGA and hear it read a draft back, and if the turn routes to a local model through Omega-MLX, the whole exchange happens without a network in sight. I wrote about the local model side in an earlier post; the short version is that Apple Silicon is absurdly good inference hardware, and speech models are small enough that on-device recognition stopped being a compromise a while ago.

The second is privacy, and I mean something concrete by that, not a vibe. People say things to a microphone they would hesitate to put in a text box: a client's name attached to a half-formed complaint, or a health question they are still working up the nerve to ask a doctor about. On most voice products that raw audio streams to a vendor. On OMEGA it does not, full stop. Even when a conversation routes to a cloud model because you picked one for that tier, what leaves the machine is the text of the turn, never the recording of your voice. The security page covers the full data-handling picture.

Dictation also changes how I use the rest of the product. My habit now is to ramble a messy paragraph at the mic and then say "tighten that up," because the mic is not a separate mode with a smaller brain behind it. The same agents that answer typed requests answer spoken ones, all 200+ of them, backed by the same 5,300+ bundled skills. Ask out loud for this morning's notes as a formatted document and the same pipeline that exports 47 formats does the work. The microphone is a front door, not a feature flag.

What the cloud voice modes get right

ChatGPT's voice mode is genuinely pleasant to use. The conversational quality is high, and the back-and-forth feels natural in a way that took real engineering. Replit shipped Voice Mode for its agent in June, and I think the instinct is right: describing an app out loud while it builds is a lower barrier than a blank prompt box. These are polished products made by serious teams.

The disagreement is architectural. Both route your audio through vendor servers, because both are cloud products and that is the only place their models live. For casual use the trade is fine. For work, where the things you say out loud include client names and unreleased plans, shipping raw audio off the machine is a real cost, and the interface never warns you about it. Lose your connection and the voice goes with it.

The second limit is harder to see because it feels normal: cloud voice modes end at the speaker. Voice, for them, is an input and output skin on a chat session. Neither one can touch a phone line. The assistant can tell you the dentist's number. It cannot call the dentist.

The part where it actually dials

OMEGA places real phone calls. Not simulated ones, and not a reminder nudging you to do it yourself. It dials the number and speaks with its own voice, and when the call ends you get an outcome, not a to-do item.

Here is a booking, step by step. You say, "get me a table for four Saturday, around seven, outside if they have it." OMEGA resolves the restaurant's number, then shows you a call plan before anything rings: the number it intends to dial and exactly what it is allowed to accept on your behalf. Maybe you widen the window to six-thirty through eight. You approve the plan. It dials. You can watch the live transcript scroll by or go do something else. When the call ends you get the result and the details worth keeping, like the confirmed time and the name the reservation is under.

Follow-ups work the same way. The contractor who has not answered two emails will usually answer a ringing phone. The clinic that needs you to reschedule does not do it over email at all. A surprising amount of everyday coordination still lives exclusively on the phone line, which is precisely why an assistant that ends at the speaker keeps handing the errand back to you.

The guardrails matter as much as the dialing. OMEGA identifies itself as an assistant calling on your behalf, because the person on the other end deserves to know. It stays inside the plan you approved: if you authorized a reservation and the restaurant asks for a card number to hold it, that request is outside the plan, and the answer is that you'll confirm directly. It does not improvise commitments. This is the same approval-first posture OMEGA takes with computer use, where every application it can touch is granted by you, per app. And under the hood a call is planned by the same Neural-Fractal Agentic AI™ (NFA) engine as any other task: a scoped objective with explicit constraints, run by a unit that reports back up when it is done.

Where calls go wrong

Some failure modes turned out to be non-problems. Phone trees are menus, and menus are easy. Hold music does not bore software; OMEGA will wait on hold with a patience no human should have to summon.

The real failures are human. Some receptionists hang up the moment they realize an assistant is on the line, and that is their right. Some businesses have a policy against it. Voicemail is another gap: OMEGA can leave a clear message, but a message is not an outcome, and it will tell you the difference instead of pretending the errand is done.

Ambiguity is a hard boundary too. If you approved a window ending at eight and the only open table is eight forty-five, OMEGA does not guess what you would have wanted. It declines and brings the offer back to you. That is correct behavior, and it still means the table is not booked. Sometimes the second call has to be yours.

When a human should make the call

This boundary is what makes the feature trustworthy. Delegate calls whose outcome is a fact or a booking. Keep calls whose outcome is a relationship. In practice:

• Anything with emotional weight. A condolence call or a hard apology. Sending software to make that call means declining to show up.
• Negotiations. A salary conversation or a contested invoice. Tone and silence carry information in both directions, and you want a human reading them.
• Medical and legal conversations that branch. When new information mid-call changes what you want, the judgment being exercised should be yours. OMEGA can book the appointment; it should not be the one hearing results.
• Relationships with a long horizon. Your longtime client, or your kid's school. Even a flawless assistant call tells the other person something about how much the relationship is worth to you.

Most of these calls OMEGA could complete mechanically. Completing the call was never the point of them.

One brain, another surface

The reason voice earns a place inside OMEGA rather than sitting on top of it is memory. The voice surface reads the same 4-layer memory and the same Brain knowledge graph as every chat window, so "call the restaurant we booked for Dad's birthday last month and move it to Sunday" resolves against things OMEGA already knows. You do not re-explain context to the microphone. And because OMEGA is provider-portable, the model answering your spoken question can change mid-conversation without the context resetting.

OMEGA is a native macOS app for Apple Silicon, macOS 14 and up, and on-device voice ships as part of it; details are on the pricing page. My bet here is simple. Voice input without privacy is a party trick, and voice output without the ability to act is a screen reader. Put the speech on the device and a phone line under the assistant, and voice finally does work you can feel: the errand you dictated at nine is a confirmed reservation by ten past.