Which AI models do dictation apps use?

Many AI dictation apps use two models. A speech recognizer turns your voice into raw text, and an optional language model then edits that text. The speech model can be a Whisper or Parakeet-style network, or a vendor's own model, and the language model can be a cloud model from a large AI company. Vendors disclose very different amounts, and for some apps the only account is a press report. The name matters less than where each model runs and where your words go.

Every AI dictation app has a speech model, and many add a language model

The first model is the speech recognizer. It is the one that hears you, and every dictation needs it. Superwhisper's models page describes the split plainly: "A speech recognition model turns your voice into text." Then an optional language model rewrites that text in Super Mode ¹. The second model is what the apps add to clean up the output, and it is the one that can change a word you said.

The two models can come from different companies and run in different places. Superwhisper's page gives a common setup as "On-device Parakeet into cloud Claude Sonnet" ². In that setup, the speech step runs on your computer, and the rewrite is sent to a cloud service. Other combinations are possible, and each one has its own privacy and speed trade-off.

Superwhisper names its speech and language models

Superwhisper is the clearest case. Its models page lists its speech and language models, and it marks which run on the device and which run in the cloud.

The route matters as much as the model name. A cloud model sends your text to a server, and an on-device model does not, so read both when you check an app.

Wispr Flow's models were reported, not documented

Wispr Flow is less specific. In a June 2025 report, 9to5Mac wrote that Wispr Flow's intelligence is based on a mix of Meta's Llama and OpenAI's Whisper models ³. That is a press account of an interview, and the report does not say which model handles the speech step. The company also told the reporter that "as they expand their feature set, they may bring in other model providers" ⁴.

So the honest answer for Wispr Flow is that the models are not fully documented in what we could find, and they may change. Wispr's privacy page says dictated audio goes to its servers over an encrypted connection ⁵, so you know the route even where the model is not named.

Some big-company apps name a category, not a model

OpenAI's ChatGPT dictation records your audio and sends it "to our models to be transcribed," and the help page lets you edit the text before you send it ⁶. The FAQ does not name the speech model. So if you want to know which one hears you, the answer is not in that page.

Microsoft's voice typing is more specific about its method and less about its model. Microsoft says fluid dictation is "powered by on-device small language models (SLMs) to ensure fast and private processing" ⁷. That names a category. The help page does not name the specific model in that sentence.

Apple describes the model for the newest iPhone dictation only as on-device. With Apple Intelligence in iOS 27, dictation uses an on-device model to improve spelling, punctuation and capitalization ⁸. That is available in English on the iPhone models Apple names.

The names matter less than the route your words take

For most people, the question behind "which model" is "where does my text go?" Three answers cover most apps:

• Fully on the device. The speech and language models both run locally, so your audio and text stay on the computer. Local models can differ from cloud ones in quality and speed, depending on the hardware.
• Speech on the device, rewrite in the cloud. The recognizer is local, but the text is sent to a cloud language model for the edit. Superwhisper's Parakeet and Claude setup is an example.
• Both in the cloud. Audio goes to a server, and so does the rewrite. Wispr Flow's stated design is an example of this route.

It depends on the app you pick and how you configure it. A raw mode that skips the rewrite removes the language model from the path, so fewer parties see the text. The question on whether AI dictation can run on your own computer covers that setup in full.

What to do

1. Find the speech model and the language model in each app's settings or documentation. If a model is not named, write that down as unknown.
2. Check the audio route: local, a proxy, or the vendor's servers. The privacy page or the model list usually says.
3. Check whether the language model can be turned off, and whether you can use a raw mode for sensitive text.
4. Treat a press report as a lead, not a fact. Where a vendor's documentation and a report disagree, go with the vendor's current page, and say so when you cite it.
5. If you need a single answer, ask the vendor in writing which speech model and language model process your audio and text, and where.

Yap, which we make, is an on-device dictation app for Mac and Windows, so the speech step runs on your computer. Our download page and pricing page cover what we offer. For the question of whether an AI step changes your words, read the guide to AI cleanup for dictation, and for the broader picture, the guide to AI dictation and the our overview of voice and AI. For the term, see the glossary entry on on-device speech recognition and the glossary entry on cloud dictation. The question on whether AI dictation reads your screen covers the context that these models receive.