Technology & privacy
What local speech-to-text is, mechanically
A model file on your disk turns audio into text without the audio leaving the machine. Here is how that works, why it became possible, and where the limits genuinely are.
Last updated
What actually happens to your voice
Five steps, none of which require a background in machine learning to follow.
- Sound becomes numbers Your microphone samples air pressure thousands of times a second. That is all audio is: a long list of numbers describing a wobbling membrane.
- Numbers become a picture of the sound The stream is converted into a representation of which frequencies were present at which moments — essentially a picture of the sound over time. Speech models work on this rather than on raw samples.
- Silence is trimmed Voice activity detection finds where speech starts and stops so the model is not asked to transcribe room tone. This is also why the first and last words are the ones that go missing if you speak too early or trail off.
- The model proposes words This is the part that is genuinely a neural network. It has been trained on an enormous quantity of audio paired with its transcript, and it produces the most likely sequence of words for what it was given — using the surrounding words to decide. That is why a whole sentence is more accurate than the same words spoken one at a time.
- The words are corrected and placed Your vocabulary corrections are applied, punctuation is placed, and the text goes where your cursor is.
"Local" means step four happens on your machine. Everything else was always local; step four is the one that used to need a server.
Why it can run on a laptop now
The weights were published
Speech models used to be proprietary and rented. When OpenAI published Whisper's weights openly, high-quality recognition stopped being a service and became a file. Anyone could download it and run it. Parakeet and other families followed.
That single change is why a dozen local Mac dictation apps appeared within about eighteen months. The hardest part of building one had been removed.
The hardware stopped being the obstacle
Running a model means moving a large number of weights through arithmetic units, repeatedly. On a traditional machine the weights live in system memory and have to be copied to the graphics card's separate memory to be processed quickly — and copying gigabytes is the slow part.
Apple Silicon's unified memory means the CPU and GPU address the same memory. There is no copy. That, plus a dedicated Neural Engine, is why a MacBook can run a speech model in a fraction of a second and an older machine cannot. See why Apple Silicon suits local speech AI.
What a model file actually is
It is worth demystifying, because "AI model" sounds like something that must live in a data centre.
| Typical for dictation | |
|---|---|
| What it is | A file of numbers — the learned weights of a neural network |
| Size | A few hundred megabytes to a couple of gigabytes |
| Where it lives | The app's data directory, usually under ~/Library/Application Support |
| How often it changes | When the app ships a new one. Otherwise never |
| What it contains | No audio, no transcripts, nothing about you. It is a trained artefact |
| What it needs to run | RAM to hold it and compute to push data through it |
How to tell if an app is really local
Three signatures, none of which can be faked by a marketing page:
- A first-run model download. A few hundred megabytes to a couple of gigabytes. An app with nothing to download is not running a model locally.
- Model files on disk. Look in the app's folder under Application Support.
- It works with the network off. The definitive test. Full method at does dictation upload your audio.
The honest limits
- Model size is bounded by your machine. A data centre runs bigger models, and on genuinely difficult audio that shows.
- Fewer languages. Local apps ship what fits; cloud services ship everything.
- Cold starts. A model has to be loaded into memory. The first dictation after an idle period pays for that, unless the app holds a gigabyte resident all day — which is its own cost on a laptop.
- It uses your battery. Modest on Apple Silicon, not zero.
- Disk space. A couple of gigabytes, permanently.
Questions
Does a local speech model learn my voice?
No. The model's weights are fixed when it ships and do not change from use. What adapts is the vocabulary layer around it, which biases towards spellings you have taught it — and that is stored locally.
How much disk space does it need?
Typically a few hundred megabytes to a couple of gigabytes for a dictation-grade model, downloaded once and reused.
Does it need a GPU?
Not a separate one. On Apple Silicon the integrated GPU and Neural Engine do the work, and unified memory removes the copying that makes this slow on other architectures. That is why Apple Silicon is a requirement for several local dictation apps.
Is local speech-to-text the same as offline?
Nearly, and the distinction is worth keeping. Local means the recognition runs on your machine. Offline means the whole app works with no connection, which is stricter — an app can be local and still refuse to start without a licence check.
Can I run speech-to-text locally without an app?
Yes — the open model families can be run from the command line, and people do. An app adds the parts that make it usable: silence trimming, getting text into other applications reliably, custom vocabulary, and managing memory so a model is not resident all day.