OpenAI Released the Brain. Every Voice App Runs on the Same One.

There is a version of building software where you have a secret. A proprietary algorithm, a dataset nobody else has, a model you trained that your competitors cannot match. For a long time, speech recognition was exactly that kind of business — accuracy was the secret, and the companies with the best accuracy won.

Then OpenAI published Whisper, and the secret stopped being a secret. We have built our company in the world that exists after that moment, and we think it is worth explaining what that world actually looks like — because if you are choosing a Whisper app for your Mac, the most important things to compare are not the things the marketing pages shout about.

The day OpenAI gave away the brain

On 21 September 2022, OpenAI released Whisper as open-source software — code and trained model weights, under the permissive MIT licence — alongside a research paper with the very academic title "Robust Speech Recognition via Large-Scale Weak Supervision." The plain-language version: OpenAI built a speech recognition system that was genuinely excellent, and then let anyone on earth download it, run it, and build a business on it, for nothing.

It is hard to overstate how unusual that was. Speech recognition had been a fortress business for decades — think of the long reign of the company behind Dragon. Accuracy was guarded, licensed, and expensive. Whisper did not just compete with that model; it removed the reason the fortress existed. Within months, an ecosystem of speech apps sprang up, and the thing they all had in common was that none of them had to build the hard part.

What Whisper actually is — and what "open source" means here

Whisper is a neural network trained to turn audio into text. What made it special was the training data: 680,000 hours of audio scraped from the open web — multilingual, multi-domain, gloriously messy. Podcasts, lectures, interviews, videos, phone-quality recordings. Most speech models before it were trained on cleaner, narrower datasets and were brittle the moment real life intervened. Whisper heard the mess during training, so it handles the mess at runtime: accents, background noise, crosstalk, technical vocabulary.

"Open source" here is doing a lot of work, so it is worth being precise. OpenAI released the model weights — the actual trained network — not just the code to run it. That distinction matters. It means a small team can take Whisper, run it entirely on a user's own machine, and never depend on OpenAI for anything. No API key, no per-minute fee, no rate limit, no terms of service that can change. The model became infrastructure, the way a font or a programming language is infrastructure. You build on it; you do not pay rent on it.

Three years on: everyone runs the same model

Look at the local-speech category on the Mac today and you will find a striking sameness under the hood. MacWhisper, SuperWhisper, EmberType, Aiko, and a long tail of smaller apps — the engine is Whisper, in one variant or another. A great deal of that is owed to one more open-source project: whisper.cpp, a high-performance reimplementation of Whisper in C/C++ by Georgi Gerganov, tuned hard for Apple Silicon. whisper.cpp is the reason running the large model locally on a laptop went from "technically possible" to "actually pleasant," and it is the quiet engine inside a large share of the apps just listed.

So picture the category honestly. A dozen apps, a shared free brain, a shared free runtime to make that brain fast. If you only looked at the core transcription, you would struggle to tell most of them apart — and that is not a knock on any of them. It is the natural endpoint of a commoditised component. The interesting question is what happens to a market when its hardest part becomes its most equal part.

Why accuracy stopped being a feature

Here is the part that took us a while to accept, as builders. We are proud of EmberType's accuracy. But we cannot honestly tell you that our accuracy is dramatically better than a well-configured competitor's, because we are all feeding audio to close relatives of the same model. When everyone runs Whisper large, everyone lands in the same neighbourhood. The differences that remain are real but small — a percentage point here, a bit better punctuation there — and they are mostly downstream of configuration, not of some secret model.

This is why we get a little uneasy when we see voice apps competing on accuracy claims as if it were 2018. "The most accurate dictation app" is, in 2026, close to a meaningless sentence, because the accuracy ceiling is a shared public asset. It is a bit like a brand of bottled water advertising that it contains hydrogen and oxygen. True; not a differentiator.

Accuracy has not stopped mattering. It has stopped being a moat. Those are different statements. A voice app still has to be accurate — but accuracy is now the price of entry, not the prize. The prize moved.

The test that proves the convergence

You do not have to take our word for any of this. If you have a Mac and twenty minutes, you can watch the convergence happen.

Install two or three Whisper-based apps. Make sure each is set to a comparable model size — a medium or large Whisper model, not the tiny one. Dictate the same paragraph into each: ideally something with a couple of proper nouns and some natural punctuation. Then compare the raw transcripts.

What you will almost certainly find is that the transcription quality is close enough that you would not reliably pick a winner blind. Where you will feel a sharp difference is everything else: how fast the text appeared, whether it landed in the app you were actually typing in, whether you had to open a separate window, whether it asked you to log in, whether it felt instant or laggy. You set out to test accuracy and you end up testing product design. That is the whole essay in one experiment.

So where is the moat now?

If the model is shared, the competition has to live somewhere else. After three years of building in this category, we would put the real differences into three buckets — and we would encourage you to evaluate any voice app, ours included, against these and not against accuracy adjectives.

• Latency — how fast the words appear after you speak, and how consistent that speed is.
• Architecture — where your audio goes, what is stored, what the app depends on to function.
• Integration — how well the result lands in the actual app you are working in, with the formatting and corrections you needed.

None of those three is glamorous. None of them is a model. All three are the result of unglamorous engineering choices, and all three are where a voice app is actually won or lost in 2026.

Latency: the moat you can feel

Dictation is a real-time act. You are speaking at the speed of thought and you want the text to keep up. The gap between finishing a sentence and seeing it appear is the single most felt property of a dictation tool — more felt, day to day, than a small accuracy difference, because you experience it on every utterance.

And latency is genuinely hard to get right, which is exactly what makes it a moat. It depends on which Whisper variant you run, how you have it quantised, how you stream audio into it, how you use the Mac's GPU and Neural Engine, how you chunk long speech, and how you overlap recognition with the user still talking. A cloud-based app adds the network round-trip on top of all of that. Two apps running "the same model" can feel a generation apart because one team sweated the pipeline and the other shipped the reference implementation. The model is free; making it feel instant is not.

Architecture: the moat you can't see (but should ask about)

This is the one we care most about, and the one that is hardest for a user to evaluate, because architecture is invisible. Two apps can produce an identical transcript while doing completely different things with your audio on the way to that transcript.

One app runs Whisper entirely on your Mac: the audio is captured, transcribed locally, and discarded, and nothing ever touches a network. Another app sends your audio to a server, transcribes it in the cloud, and sends the text back — possibly faster, possibly using a bigger model, but your voice has now left your machine and you are trusting a privacy policy. A third does a hybrid of the two. From the outside, looking only at the transcript, you cannot tell which is which. You have to ask.

We made the architecture choice early and we made it absolute: EmberType runs Whisper 100% on-device, with no cloud fallback, no account, and no telemetry. That decision cost us things — it ruled out the easy path of renting a big cloud model — but it is the one part of our app that a competitor cannot copy by also downloading Whisper. Anyone can have the model. Not everyone is willing to give up the cloud. Architecture is a moat precisely because it is a commitment, not a feature, and commitments are expensive in a way that features are not.

So when you evaluate a voice app, ask the question the transcript cannot answer for you: where does my audio go, and what is kept? If the answer is "nowhere, and nothing," you have found the kind of app whose privacy does not depend on anyone keeping a promise.

Integration: the moat that is pure unglamorous work

The third moat is the least exciting and possibly the most decisive. Whisper hands you a block of text. It does not put that text into your email, your code editor, your Slack message, or your Google Doc. It does not know that "function" should not be capitalised in your editor but "Monday" should be in your calendar. It does not strip the "um" you said while thinking. It does not know that when you said "new paragraph" you meant a command, not a phrase to transcribe. Everything between Whisper's raw output and text usefully landing where you needed it is work the app has to do, and Whisper does none of it.

That gap is enormous, and it is pure product engineering. Real-time dictation that types into any app is a different and harder product than batch transcription of a file, even though both can sit on the same Whisper model. Contextual cleanup, custom vocabulary, the keyboard shortcut that works everywhere, the cursor landing in the right place — that is months of unglamorous engineering that never shows up in an accuracy benchmark and entirely determines whether the tool is pleasant to use. The model is the easy 20%. The integration is the hard 80%.

Whisper's honest limits — and what comes after Whisper

We would be doing exactly the thing we criticised — overclaiming — if we pretended Whisper is flawless. It is not. The tiny and base model sizes are genuinely weak and will frustrate you; they exist for speed-constrained situations, not for serious dictation. Whisper can "hallucinate" plausible-sounding text during long silences. It does not natively diarise — it will not, on its own, tell you who said which sentence. And like any model trained on web data, it is stronger in some languages than others.

It is also no longer alone. As of 2026, other open speech models from other labs have appeared and compete with Whisper on speed and accuracy — you may have seen comparisons fly past. We watch that landscape closely, and here is the thing: a stronger or faster open model appearing does not threaten the argument of this essay. It confirms it. Each new open model commoditises the brain a little further. The model layer is a rising tide that lifts every boat in the category equally. Which means, again, that the model is not the moat — and whatever app you choose, you should choose it for the three things that are.

How to choose a Whisper app for Mac

Put all of this together and the buying decision gets simpler, not harder. Ignore accuracy adjectives. Run the twenty-minute test above to confirm the transcription is in the same neighbourhood — it will be — and then decide on the three real axes.

• Match the tool to the job. Real-time dictation into any app, batch transcription of recordings, and meeting capture are three different products. An app optimised for one is rarely the best at another. Pick for your actual primary use.
• Decide your architecture line. If your work touches anything sensitive — or you simply prefer it — choose a fully on-device app and confirm it has no cloud fallback. If you do not mind the cloud, you have more options. Just decide on purpose.
• Test the latency yourself. It is the property you will feel a hundred times a day. A free trial tells you in five minutes what a spec sheet never will.
• Check the pricing shape. A model that costs nothing to license does not require a subscription to wrap around it. Plenty of good Whisper apps are one-time purchases. A recurring fee should buy you a recurring service, not just access to a free model.

That is the honest buyer's guide. It does not end with "and therefore buy ours." It ends with: know what you are actually comparing, because the marketing in this category is still fighting the last war.

What to do this week

If you dictate on a Mac, do the twenty-minute test. Install two or three Whisper apps, set them to comparable models, and dictate the same paragraph into each. Watch how little the transcripts differ and how much everything else does. You will never read a voice-app marketing page the same way again — and that clarity is worth more than any single download.

If you want one of those apps to be the on-device, no-cloud, no-subscription end of the spectrum, EmberType is free for seven days. We are not going to claim it transcribes better than the model everyone shares — that would be the exact dishonesty this essay is about. We will claim we spent our years on the parts that are not the model: the latency, the architecture, and the integration. Those are the parts you can actually feel. Go feel them.

Frequently Asked Questions

---