AI dictation

AI dictation is dictation with a language model added after speech recognition. The recognizer turns your voice into a raw transcript, and the model then edits that transcript. Depending on the app, it removes filler words such as "um", adds punctuation, fixes capitalization or rewrites a sentence to suit the app you are typing into. In many products the model works on the text the recognizer produced, so the edit happens in a second step.

A language model can fix words the recognizer missed

A recognizer chooses among candidate words. A language model can do more. In the HyPoradise benchmark, a large language model can correct words that are missing from the recognizer's N-best list, because it generates text rather than only choosing among candidates ¹. That benchmark measures recognizer output in research, not any commercial app, so it shows what the method can do and not how often a product uses it.

An example of AI editing

Suppose the raw transcript reads: "so um I think we should move the meeting to thursday comma no wait friday period." An AI editing step might return "I think we should move the meeting to Friday." This is an illustration of the kind of edit, not a measured output. The app's settings decide how far the model goes, and a stricter setting may keep "Thursday" and the correction in place.

Plain dictation and AI dictation are often separate settings

Many products let you turn AI processing on or off. Superwhisper's Free plan, for example, uses local Whisper models without AI processing, and the company says nothing dictated leaves the device ². Its paid plan adds cloud models and AI processing on top.

The setting matters for privacy. HEX's README says that when post-processing with OpenCode is turned on, the app sends the transcript, the foreground application and the Brave website hostname to the configured model provider ³. Check whether the rewrite is on before you dictate sensitive text.

Why AI editing matters: it targets the slowest part of dictation

Correcting dictated text is the slow part. In a 2006 study, correcting errors took up most of the task time when people dictated with a large-vocabulary recognizer ⁴. AI editing aims to cut that work by removing filler and fixing punctuation for you.

It can also change meaning. A model that tidies a sentence may alter a figure, a name or a legal or medical phrase. Read the output before you send it, and keep a plain-dictation option for text where every word has to match what you said.

Where to read more, and related terms

For a comparison of the major apps, read AI dictation apps compared. For how cleanup works and what it costs in practice, see AI dictation cleanup, and for setups without AI, see speech-to-text without AI. The question whether there is a dictation app without AI covers the same choice from the other side.

AI dictation is a kind of dictation software that often runs in the cloud, though some tools keep the editing model on the device. A word list helps the model as much as the recognizer, so see custom vocabulary. Some products let you supply your own model key, described in bring your own key. Many AI dictation apps also offer a free tier with limits, so check those before you rely on one.