In August we measured whether speech recognition can run on the phone itself. The result was unambiguous: no. We wrote the numbers down, gave reasons for the decision and put recognition in the cloud.
The measurement was right. The conclusion was wrong. We had asked the wrong question — and the answer had been on the device the whole time.
What we had measured
The setup: the test device was a Nokia 7.2 — Snapdragon 660, built in 2019, Android 11. Deliberately a mid-range phone from years back: if it holds up there, it holds up everywhere. Measured on 21 and 25 August 2026.
We had packed Whisper into the app — the open model from OpenAI, in the size “small”, 190 MB. The quality was good: with the vocabulary it recognised “L-Alanine” and “L-Carnitine” just as reliably as the cloud did. The smaller model “base”, on the other hand, made the one mistake that has no place in the diary of a health app:
Spoken: “half a gram” — recognised: “½ kg”.
A wrong name is something you notice. A wrong factor of 1,000 is not — at any rate not in the morning, not in a hurry.
Speed was where it ended. On a Snapdragon 660, “small” took around 30 seconds per utterance, almost regardless of its length, because Whisper computes internally in windows of 30 seconds. The same file via the cloud: 1.4 seconds. The user who tried it put it more briefly:
“So bad that you can basically type faster than wait for the translation.”
On top of that, 190 MB in every download. After the removal, the directory holding the bundled files shrank from 182 MB to 968 KB.
The mistake in our thinking
All of these numbers still hold. They just answer a narrower question than the one we thought we had asked.
What we had measured was: can a model we ship ourselves compute fast enough on a mid-range Android? What we had understood it as was: can a phone do speech recognition?
Those are not the same question. Every modern phone can do speech recognition — it does it every day when somebody taps the microphone on the keyboard. That recognition is part of the operating system, it is matched to the device’s hardware, it is already installed, and it costs nothing. We had built a model of our own alongside it and measured it against the cloud instead of against what was already there.
The old article even wrote down when we would measure again. One of the three triggers read, word for word: “or the operating systems come with recognition of their own”. They came with it. We had simply not looked.
What forced us to look
Not curiosity, but a letter. We had asked the recognition service in writing whether it may carry health data under Article 9 GDPR. The answer was no — the provider’s certification covers servers, storage and network, not the recognition service. The same at a second provider.
A third provider would have been the obvious next step. Instead we asked the question again from the start — and the answer was in the operating system.
How it works now
The app uses the operating system’s recognition and expressly insists that it works on the device: on iPhones Apple’s Speech framework, on Android the on-device recogniser. It only starts if the system confirms that it understands the selected language without a network.
Three things have got better as a result, and without anything to set against them:
- The recording no longer leaves the phone. No service provider, no contract, no transfer, no consent.
- The text is there while you speak. Recognition delivers interim results as it goes instead of one answer at the end — the cards take shape while you are still talking.
- No model in the package. The 190 MB are not coming back.
There is no new stopwatch measurement like the one in August yet, and we are not inventing one. What you can see without a stopwatch: there is no waiting for an answer, because there is no answer that could be waited for.
What it costs
The honest part, and this time it is shorter than last time, but it is there.
Not every phone can do it. On iPhones recognition has been on board for years. On Android it takes version 13 or newer and an installed language pack for recognition without a network. If both are missing, there is no speech recognition in RecTake — then you type. The Nokia 7.2 that was measured on in August is one of those phones itself.
We are not building a silent fallback to a server for that. An app that transmits after all when in doubt and does not say so is worse than one that lacks a feature.
And we have fewer knobs to turn. With a model of our own you could choose the size, build the prompt, adjust the parameters. With the operating system’s recogniser there is exactly one lever: the vocabulary the app passes in as context. It decides whether “Elalanin” becomes “L-Alanine”. That has made it more important, not less.
What we take away from it
The measurement in August was clean, the device was deliberately a poor one, the numbers were real. The mistake sat one level above: in the assumption that “on the phone” and “our model on the phone” are the same thing.
So we are not simply making the old text disappear, but writing down what was wrong with it. Where the recording goes today is in an article of its own: Where your voice recording goes — the answer is one word long by now.