What actually limits an old phone
Three things, and they are not equally important.
Memory is the real constraint. A speech model has to be held in memory while it runs. On a phone with plenty, that is invisible. On one that is already juggling a browser and a messaging app, the system has to push something else out first, and that shows up as a delay before your text appears, or as the keyboard restarting.
Processor speed decides the wait. An older chip does the same arithmetic more slowly, so the pause after you stop speaking is longer. Annoying, but not fatal.
Storage decides whether you can install anything at all. A model file has to fit, and older phones are usually fuller.
Notably absent from that list: the age of the phone as such. A five-year-old device with 4 GB of memory and free storage will do better than a three-year-old budget model with 2 GB and a full disc.
The adjustment that changes everything
Use a smaller model.
Almost every complaint about dictation being slow on older hardware is solved by this one change, and it is usually the setting nobody touches because larger sounds better.
LocalType ships three sizes. The 60 MB option is built precisely for this case: it returns text fastest and needs the least memory to run. The 190 MB middle option is what the app suggests for most phones. The 539 MB one is for devices with room to spare.
On a phone that is struggling, the smallest model does not merely feel faster, it can be more accurate in practice. A model that cannot load comfortably spends its time being evicted and reloaded, and none of that time goes on your speech.
For scale, on the 4 GB test handset the middle option turned 6.9 seconds of speech into text in 4.1 seconds, while the largest takes roughly twice as long as you spent speaking. On a weaker device those gaps widen, and the choice exists for exactly that reason.
Where the accuracy actually differs
A smaller model costs you accuracy, and the cost is not spread evenly.
The difference is smallest where most dictation happens: a clear voice, close to the microphone, in a reasonably quiet room, speaking ordinary sentences. Under those conditions the smallest model is perfectly usable, and for messages and notes most people would struggle to tell you which one they were running.
The difference is largest exactly where things are already hard: strong accents, background noise, several people talking, fast speech, unusual names. Those are the cases the extra capacity was built for.
So the question is not "how much accuracy am I giving up" in the abstract. It is "do I dictate in a noisy environment". If not, the smallest model costs you very little.
Practical things that help beyond the model
Free some storage first. A phone with almost nothing left performs badly at everything, and installation briefly needs room for two copies of the model.
Close what you are not using. Memory pressure is the main cause of slowness, and an old phone has less to spare.
Dictate in single sentences, not in whole paragraphs. Shorter bursts finish sooner and sit in memory for less time.
Get closer to the microphone. Costs nothing, and improving the input helps a small model more than it helps a large one.
Add your own vocabulary. Names and jargon fail on every model size, so fixing them by hand once is worth more on old hardware than anywhere else.
Where the line genuinely is
LocalType runs on Android 8.0 and newer, which reaches back a long way and covers most phones still in daily use.
That is the software line. In practice the limit is memory, and it is not a single number, because it depends on what else you run. A device that struggles with the largest model may be entirely comfortable with the smallest. Three sizes exist because of that, and one size would not do the job.
If dictation is slow after switching to the smallest model, and storage is free, and nothing else is running, then the honest answer is that the phone is at its limit. On-device processing has a boundary somewhere, and no piece of software gets round it. Saying so plainly seems better to us than pretending otherwise.
Why local is often better on old hardware, not worse
Cloud dictation on an old phone is not free either. It keeps the radio active, which is one of the hungriest components in the device, and on an older battery that matters. It also fails on poor connections, and older phones frequently have worse antennas and are used by people on cheaper plans.
Local recognition removes the network from the equation entirely. That work moves to the processor, and you can size it down to match what the phone can do. That control is not available with a cloud service, where you take whatever latency the connection gives you.
Tablets and hand-me-downs
Two situations that are common and rarely written about.
An older tablet used as a second screen or a reading device usually has the same profile as an older phone: modest memory, plenty of storage, and no pressure to be fast. Dictation works there on the same terms, and a tablet is often a more comfortable device to dictate into because it sits on a table and does not have to be held.
A phone handed down to a child or a parent is the other frequent case. Here the smallest model is almost always right, since the alternative is a device that feels slow at everything, and the person using it will not be adjusting settings to investigate why.
For both, do the setup on the device before handing it over. Download the model on Wi-Fi, confirm it works, and add the handful of names that will be dictated constantly. What comes back to you is a device that behaves consistently, not one that generates support calls.
What you get regardless of the phone
None of the rest depends on the hardware.
Recognition happens on the device from a model file no other app can open. Internet access is used once, to collect the model you picked, and never again after that, so tunnels, aeroplanes and dead spots behave the same as home. There is no registration and no advertising. Nothing inside the product records your usage, and no audio is stored anywhere. Android's cloud backup is deliberately not given the app's data. Over a password field the microphone is not offered.
Eighteen languages come from a single multilingual model, identified from the recording, so an old phone gets the same language coverage as a new one.
Old phone, small model
Free some space, close what you are not using, and expect a short wait rather than none.
For messages, notes and email in a quiet room, that combination works considerably better than the general tone of writing about on-device processing would lead you to believe.
And the sensible move is to try it, not reason about it. Phones vary more than their specifications suggest, and ten minutes with the smallest model tells you more than any comparison of memory figures.