# AI Dictation: What the AI Actually Does to Your Words

> AI dictation pairs speech recognition with a language model that cleans your words. See what it changes, what it shouldn't, and how Lyre runs it locally.

Canonical: https://uselyre.com/blog/ai-dictation.md

[Dictation](https://uselyre.com/blog/category/dictation)

Nemo · November 16, 2026 · 5 min read

AI dictation is speech to text with a second step: after a speech model transcribes what you said, a language model rewrites the transcript into readable text. It removes filler words, false starts and repetitions, fixes punctuation, and can even translate. Done well, it keeps your meaning and voice. Done badly, it puts words in your mouth.

Here’s what each layer does, what to watch for, and how to judge an AI dictation tool before you trust it with your writing.

## What are the two layers of AI dictation?

Most modern AI dictation tools have two separate models working in sequence.

| Layer | Job | Example model |
|---|---|---|
| Speech recognition | Turns audio into a word-for-word transcript | Whisper |
| Language model | Turns the transcript into clean text | Qwen3, GPT, Claude, Llama and others |

The first layer decides whether the words are right. The second decides whether the text reads well. They fail in different ways, which is why it helps to understand both.

## What does the speech recognition layer do?

It listens and writes down what it hears. Whisper, the open model from OpenAI that many dictation apps use, is trained on a huge amount of multilingual audio, so it handles accents, background noise and natural speech better than older systems. It also adds basic punctuation.

What it doesn’t do is judge. If you say “so, um, I think, I think we should ship on Friday”, a good transcription gives you exactly that. Accurate, and not something you’d want to paste into an email.

## What does the language model layer do?

The language model reads the transcript and produces the version you meant to write. A well-behaved cleanup step:

- removes filler words such as “um”, “uh”, “like” and “you know”
- drops false starts (“I think we should, actually, let’s…”)
- collapses repetitions (“I think, I think”)
- fixes punctuation and capitalization
- keeps your vocabulary, your tone and your point

Optionally, it can translate: speak Portuguese or Spanish, get English text. That’s useful if you think faster in one language but write in another.

The [filler words](https://uselyre.com/blog/filler-words) guide shows before-and-after examples of this step.

## What should AI dictation not do?

This is where tools differ the most. The risk with a language model is that it “improves” too much. Watch for:

- **Added facts.** It should never insert a date, number or name you didn’t say.
- **Changed meaning.** “We might ship Friday” must not become “We will ship Friday”.
- **A new voice.** If you write casually, the result shouldn’t sound like a press release.
- **Lost hedges and nuance.** “Probably” and “I’m not sure” matter.

A good test: dictate something slightly uncertain and opinionated, then compare. If the output sounds like you, minus the stumbles, the cleanup is doing its job.

## Where does the processing happen?

This matters as much as quality. Many AI dictation apps send your audio, your transcript or both to cloud servers, because large language models used to be too heavy for a laptop. That’s no longer true for this task. A mid-sized open model runs comfortably on Apple Silicon.

Lyre is a macOS menu bar app that does both layers on your Mac: whisper.cpp transcribes your speech, and Qwen3 running in a bundled llama.cpp server cleans it up, so no audio or text ever leaves the machine.

The tradeoffs of local versus cloud:

|  | Cloud AI dictation | Local AI dictation (Lyre) |
|---|---|---|
| Your audio leaves the Mac | Yes, to the vendor | No |
| Works offline | Usually not | Yes, after setup |
| First-run download | Small | Large (about 1.6 GB plus 2.5 to 5 GB) |
| Hardware needs | Almost any machine | Apple Silicon, macOS 13+ |
| Speed | Depends on network | Depends on your Mac |

If keeping words on your own machine is the priority, our guide to [speech to text offline](https://uselyre.com/blog/speech-to-text-offline) goes deeper.

## How does Lyre use AI dictation day to day?

You hover the notch, click the mic, speak and confirm. Whisper Large v3 Turbo (recommended, about 1.6 GB; smaller models available) produces the transcript, the local model polishes it, and the result goes to your clipboard with the message “Copied to clipboard, ready to paste”. You paste with Cmd+V into any app.

You can switch the polish step off if you want raw transcripts, for example when quoting someone or keeping notes exactly as spoken.

Honest limits: it pastes rather than types into the focused app, there’s no global keyboard shortcut yet, and text doesn’t stream while you talk. It runs only on Apple Silicon Macs. Pricing is $5 a month or $39 a year with a 7-day refund window and no free trial.

## How do you choose an AI dictation tool?

Ask these questions before you commit:

1. **Where does my audio go?** Look for a clear answer, not a vague privacy page.
2. **Can I turn the AI step off?** Sometimes you want exact words.
3. **Does it keep my tone?** Test with your own writing style.
4. **Does it work offline?** Planes, trains and bad Wi-Fi happen.
5. **What does it cost over a year?** Compare monthly and yearly prices.

For a broader map of Mac options, start with [Mac dictation: the complete guide](https://uselyre.com/blog/mac-dictation).

## FAQ

### Is AI dictation more accurate than regular dictation?

The transcription layer is similar in kind. The AI cleanup doesn’t make recognition more accurate, it makes the text more readable. Good speech recognition, such as Whisper, is still what gets the words right.

### Can AI dictation change what I meant?

A poorly tuned one can. Lyre is instructed to keep your meaning and tone and only remove filler, false starts and repetitions, but you should always read before sending.

### Does AI dictation need the internet?

Not necessarily. Lyre runs both models locally and works offline after the first-run download.

### Can AI dictation translate?

Yes, if the language model is asked to. In Lyre you can speak, for example, Portuguese or Spanish and get English text.

### Is my voice used to train models?

With Lyre, nothing is sent anywhere, so there’s nothing to train on. With cloud tools, read the vendor’s data policy.

[Download Lyre](https://uselyre.com/download) to try local AI dictation on your Mac.

Dictation for Mac · on-device

Lyre turns your voice into clean text, right from the notch.

Transcribed and polished on your Mac: no filler words, no cloud, and it can translate as you speak. $5/month, 7-day refund.

[See how Lyre works](https://uselyre.com/#how-it-works)

**About this article.** Written by Nemo, who builds Lyre. Facts about other apps and platforms were checked on the date above and can change. Spotted something wrong or out of date? [Tell us](mailto:support@uselyre.com).

More in Dictation

[Dictation](https://uselyre.com/blog/category/dictation) Nov 16, 2026 · 6 min

## [Mac Dictation: The Complete Guide for 2026](https://uselyre.com/blog/mac-dictation)

Mac dictation explained: Apple's built-in Dictation, local Whisper apps like Lyre, privacy, accuracy and which setup fits the way you write.

[Dictation](https://uselyre.com/blog/category/dictation) Nov 16, 2026 · 5 min

## [How to Dictate on Mac: Built-in Dictation and Beyond](https://uselyre.com/blog/how-to-dictate-on-mac)

How to dictate on Mac step by step: turn on Apple Dictation, use voice punctuation, and when a local Whisper app like Lyre gives cleaner text.

[Comparisons](https://uselyre.com/blog/category/comparisons) Nov 16, 2026 · 5 min

## [Wispr Flow Alternative: A Private, Local Option for Mac](https://uselyre.com/blog/wispr-flow-alternative)

Looking for a Wispr Flow alternative? Compare cloud and local dictation on Mac, and see when Lyre, a fully offline notch app, is the better fit.
