AI Dictation: How AI Voice Typing Works

AI dictation is two steps. A neural speech recognizer turns your voice into a raw transcript. In many newer apps, a large language model then edits that transcript into finished text: it removes filler words, adds punctuation and paragraphs, applies your spoken corrections and formats the result for the app you are typing into. The first step runs in every dictation product. The second is what separates the newer AI apps from classic dictation, and it is also the step that adds delay, sends more of your words to a server and can change what you said. For whether the term "AI" fits, see is speech to text AI?. For the wider subject, see the voice and AI overview.

AI dictation is speech recognition plus an optional language-model edit

Think of AI dictation as a pipeline with one required stage and one optional stage. The diagram shows the path, and the table compares the two stages.

Classic dictation takes the short path: the raw transcript is typed as it comes. Some apps let you choose that path. Superwhisper's documentation says a dictation can skip its language model stage, and its Voice to Text mode outputs raw transcribed text with no language model involved ¹.

The rest of this guide takes the stages in order: the recognizer, the language model, context, and then the costs. For the apps themselves, see the comparison of AI dictation apps.

The recognizer is a neural network trained on 680,000 hours of speech

The recognizer does not look up sounds in a fixed rulebook. It is a neural network that learns the mapping from sound to text from recorded speech paired with transcripts. OpenAI's Whisper, one of the best-known open models, was trained on 680,000 hours of multilingual and multitask audio with transcripts ².

Apple's SpeechAnalyzer API, which apps can use for their own transcription, is one example. Apple says that transcription with it is entirely on the device, although the language models must first be downloaded ³.

Recognizers work in one of two modes. A batch model waits for a whole phrase or recording before it returns text. A streaming model returns partial guesses as you speak and revises them as more audio arrives. That is why dictation apps can show words while you are still talking. The trade-off between the two modes returns in the section on delay below.

On a 2025 test episode, Whisper made fewer errors than Apple's engine

The most direct comparison in the sources reviewed for this guide runs several engines on the same audio. In a 2025 test by 9to5Mac, one 7-minute-31-second episode had a word error rate of 10.3 percent with Apple's speech API ⁴. Whisper Large V3 Turbo had 1.5 percent on the same episode ⁵. That makes Apple's error rate about 6.9 times Whisper's ⁶.

This is one episode, so it is a data point, not a benchmark. The author also reached Apple's speech API through an open-source command-line tool, not through the Dictation key ⁷. A second comparison comes from a 2025 neurosurgery pilot. Across all errors, a Whisper-based model had a mean word error rate of 1.75 percent ⁸, and Dragon Medical One had 1.54 percent ⁹. The difference was not statistically significant ¹⁰. That pilot compares a Whisper-based model with a medical dictation product, not with the built-in tools most people use.

So the honest answer to whether AI dictation is more accurate than built-in dictation is that the evidence is thin. One test favors a Whisper-based engine over Apple's. A pilot finds no significant gap between a Whisper-based model and a commercial product. For more on how these numbers are read, see how accurate speech recognition is.

The language-model pass turns a raw transcript into finished text

The language model does work the recognizer does not do on its own. It can supply a word the recognizer never ranked first. In the HyPoradise benchmark, a large language model can correct words that are missing from the recognizer's N-best list, meaning the alternative transcripts the recognizer ranked, because it generates text rather than only choosing among candidates ¹¹. The speech to text and LLMs guide covers that mechanism in more detail.

Apps also use the second stage for editing such as removing filler words and adding punctuation. Wispr Flow's home page says the app automatically removes filler words, adds punctuation and formats your writing as you dictate ¹². Microsoft says its Fluid dictation, on Copilot+ PCs, automatically corrects grammar, punctuation and filler words as you speak ¹³.

Which model does this work is not always published. A June 2025 9to5Mac report described Wispr Flow's intelligence as based on a mix of Meta's Llama and OpenAI's Whisper models ¹⁴. If a vendor does not say which model handles your text, treat that as an open question. The question on which AI models dictation apps use goes through how to check.

The second stage works from the transcript it receives. A text-only editor sees only the words the recognizer produced, so a poor transcript limits what the edit can recover. Context can help here, too: the setting in the next section lets the editor see the names and jargon already on your screen.

The second stage does not have to run in the cloud. Superwhisper's free plan uses local Whisper models without AI processing, and the company says nothing dictated leaves the device ¹⁵. The AI cleanup guide covers the editing choices in more detail.

Context lets AI dictation fit its output to the app you are in

Context is what lets the second stage tailor its output to the words around your cursor. Wispr Flow's privacy page describes its context awareness setting as one that "Reads limited text from the app ... so names and jargon come out right" ¹⁶. Turn it off, the page says, and the app "transcribes without reading anything on your screen" ¹⁷.

That is a real trade. Reading the text around the cursor helps with names and jargon, and it means the dictation app can see text you have written. Check that setting in each app before you turn it on. The question on whether AI dictation reads your screen covers the wider picture, and the privacy guide on voice to text covers what else leaves your device.

Each AI step adds delay, cloud processing and a chance of changed words

The costs come from different places. Streaming decides how fast words appear, the second stage adds a step, and editing can change meaning even when the recognizer heard every word correctly.

Streaming services show first words in under half a second, but first guesses can be less accurate

Streaming recognizers show a first guess while you are still speaking. In Artificial Analysis's 2026 streaming benchmark, Deepgram Flux had its first partial transcript 0.019 seconds after end of speech ¹⁸. ElevenLabs Scribe v2 Realtime had its first partial at 0.13 seconds ¹⁹, and AssemblyAI U3 Realtime Pro at 0.47 seconds ²⁰.

A first guess can be less accurate than the final text. OpenAI's hosted GPT Realtime Whisper model had a 7.5 percent word error rate on its first partial transcript, improving to 5.1 percent on the final transcript ²¹. Not every model gains this much. The same benchmark lists this one among the leaders at improving by waiting. That is why words on screen can change after you pause. The question on why AI dictation is slower and the guide on real-time speech recognition cover the streaming side.

A language-model pass adds a second wait before the text settles

Language model speed is often reported as time to first token: the time in seconds between sending a request to the service or system and receiving the first token of the response ²². The sources checked for this guide do not give the delay of any dictation app's second stage, so timing it yourself is the only reliable check.

The recognizer can be the slow part too. Home Assistant's documentation says Whisper takes around 8 seconds to process a voice command on a Raspberry Pi 4 and under a second on an Intel NUC ²³. The delay depends on the hardware as well as the model. The cost guide puts the money side next to these delays.

Cloud processing means your audio leaves the computer

In a cloud product, your audio is sent somewhere before any text comes back. Wispr Flow's privacy page says dictated audio goes to its servers over an encrypted connection, TLS 1.2 or later, so cloud processing is part of its design ²⁴.

On-device recognition avoids that trip. Apple's SpeechAnalyzer runs on the device, as noted above ³. In its report on the first iOS 27 beta, 9to5Mac said the Advanced Dictation Preview runs only on the on-device model, so its quality is the same with or without a data connection ²⁵. The question on whether AI dictation can run locally covers the offline options in full.

Editing can change words you meant to keep

A language model that rewrites text can also change what you meant. In a 2025 INTERSPEECH paper, Yamashita and colleagues say existing language-model correction methods rely mainly on text, neglect phonetic cues, and that this leads to over-correction ²⁶. Over-correction means changing words that were already right.

Invented content is the other risk. In a study of clinical notes that language models generated, the hallucination rate was 1.47 percent, measured against clinician-annotated sentences ²⁷. That is a different task from dictation, but it shows what a generated error looks like: fluent, plausible and not in the source.

Editing is also a choice about style. GoTranscript says edited transcription rewrites for clarity and flow, which helps readability but moves the text further from the spoken record ²⁸. The raw transcript is not always a faithful record either. A 2025 study of Whisper on dementia speech says the model tends to omit filler words such as uh and um ²⁹.

Built-in dictation is already AI, so the real choice is how much editing you want

On a Mac, Apple says general text Dictation is processed on your device and not sent to Siri servers ³⁰. On iPhone, Apple Intelligence in iOS 27 adds a pass that uses an on-device model to improve spelling, punctuation and capitalization. Apple says that is available in English on named iPhone models ³¹. Microsoft's Fluid dictation applies the same kind of editing on Copilot+ PCs ¹³.

So the useful question is how much editing you want, and where it runs. The table sets out three common options.

Four questions show which stage an app runs and where your words go

Before you choose, ask the same four questions of any AI dictation app. The answers are in its settings, its privacy page or its documentation, and each one maps to a risk covered above.

1. Editing step: can it be turned off? Superwhisper's Voice to Text mode, for example, outputs raw transcribed text with no language model involved ¹.
2. Audio route: check whether the audio is sent to servers or stays on the device, as Wispr Flow's privacy page describes for its service ²⁴.
3. Screen reading: Wispr Flow lets you turn its context setting off. Its privacy page says the app then transcribes without reading anything on your screen ¹⁷.
4. Editing model: if the vendor does not name it, treat it as unknown. A 9to5Mac report on Wispr Flow, for example, described its models as a mix of Meta's Llama and OpenAI's Whisper. That is a press account, not a vendor statement ¹⁴.

Yap, which we make, is an on-device dictation app for Mac and Windows, so it belongs with the on-device options in that comparison. For plain dictation, the built-in tools are often enough, and this guide does not argue otherwise. Before you pay for any app, read how Apple dictation works on a Mac and how voice typing works on Windows. Our download page and pricing page cover what we offer.

If you want a side-by-side look at the apps, the AI dictation apps compared guide sets out how the major apps differ. For what the terms mean, see on-device speech recognition and cloud dictation.