In 2026, "AI dictation" usually means two separate steps. First, a speech recognition model turns your audio into words. Second, and optionally, a language model cleans up those words: it removes "um", fixes punctuation, turns spoken lists into bullets, or rewrites the text in a different tone. Each step can run on your own device or on a company's servers, and they do not have to run in the same place. Knowing which step happens where tells you what an app sends, and what it might change.
Publisher note: SuprFlow publishes this guide and is one of the products discussed. Competitor details come from each company's own pages, checked on 1 October 2026. Prices and features change; confirm before you buy.
What are the two steps?
| Step 1: speech recognition | Step 2: language-model cleanup | |
|---|---|---|
| What goes in | Your recorded audio | The text from step 1, plus instructions and sometimes context |
| What comes out | A raw transcript | Edited, formatted or rewritten text |
| Typical models | Whisper, Parakeet and vendors' own models | General-purpose language models, run locally or through a provider |
| If it runs in the cloud, what is sent | Audio | Text (and any context the app adds) |
| Typical mistakes | Sound-alike words, misheard names | Fluent changes to meaning |
Superwhisper's documentation describes this split clearly: "Every dictation in Superwhisper passes through two independent stages", voice to text and then text post-processing, and "Each stage is independently configurable" between local and cloud (Superwhisper). The other apps below describe the same split in their own terms. How speech recognition works explains step 1 in more depth.
What does the AI cleanup actually do?
Vendors describe it in their own words. Apple says that on supported devices, Dictation "now captures what users say as polished text with greater precision, automatically handling capitalization, punctuation, and formatting as they speak", using an on-device model available on a Mac with M3 or later and at least 12GB of unified memory (Apple Newsroom). BetterDictation publishes its default cleanup prompt, which tells the model to remove stammers and filler words, remove false starts, correct grammar, choose a format such as a list, and "Ensure that your edits do not alter the intended message" (BetterDictation user guide).
That last instruction shows the core tension. The cleanup step is asked to change your words, but not your meaning, and a model can get that balance wrong.
Where does each step run?
As each vendor documents it, checked on 1 October 2026. This is a map of architectures, not a ranking.
| Product | Speech recognition | Cleanup step | What leaves the device for cleanup |
|---|---|---|---|
| Apple Dictation | Keyboard settings tell you whether general text dictation is processed on your device (guide) | Advanced dictation uses an on-device model on supported Macs | Not described as leaving the device |
| Wispr Flow | In the cloud; "Dictation is transcribed in the cloud" (help) | Part of the service: Flow "transcribes, formats and inserts your speech" (guide) | Audio goes to the cloud for step 1 |
| Monologue | Remote by default; optional local mode on Apple silicon Macs | "Smart formatting may still require a network connection" (docs) | In local mode, the docs do not say exactly what the formatting step receives |
| Superwhisper | Local or cloud voice model, per mode | Local or cloud language model, or none in Voice to Text mode | Text, if you choose a cloud language model |
| BetterDictation | On the Mac | Optional Pro step via OpenAI | "your transcribed text (but not your audio)" (privacy policy) |
| SuprFlow | On the Mac, with a downloaded WhisperKit model | Default: local cleanup of punctuation, capitalization and numbers; optional Cloud API or Local (Ollama) | With Cloud API, transcript text and formatting instructions, not audio, go to the provider you configure |
The same app can sit in different rows depending on settings. That is why "Is this app private?" has no single answer; "What does this setting send?" does.
What can the AI get wrong?
Speech recognition errors usually sound like what you said: "flour" for "flower". Cleanup errors can be harder to spot because the result reads well. Watch for four kinds:
- Names. A rare surname "corrected" to a common word, or a client's name spelled the way the model expects.
- Numbers. "Fifteen, not fifty" becoming "50", or a percentage turned into a different figure while tidying.
- Negations. "Did not approve" becoming "approved", or "no exceptions" appearing where you said "two exceptions".
- Rewritten meaning. A caveat dropped for brevity, a question turned into a statement, or a spoken self-correction applied to the wrong phrase.
For professional records, these are the errors that matter. What if an AI formatter changes an audit exception? walks through a correction workflow. If you need words exactly as spoken, look for a raw or verbatim setting and test it, because a label is not proof; Is SuprFlow's Voice to text mode really raw? shows how to check.
How do you check what an app sends?
Ask these questions of any AI dictation app, and look for answers on its own documentation or privacy pages:
- Where does speech recognition run in the mode you will use, and which model is it?
- Is there a cleanup step, is it on by default, and where does it run?
- Does the cleanup step receive audio or text?
- What context is added? Some apps read the active app, nearby text or the screen. Monologue, for example, documents an optional Deep context feature that uses screen capture permission (permissions and privacy).
- Which AI providers receive data, and on what terms? Look for named providers, retention, and whether data trains models. Superwhisper, for instance, says it uses providers through "API agreements that include zero-data-retention terms".
- Can you see the raw transcript next to the cleaned version, to compare?
- What else uses the network? Licence checks, updates, crash reports and telemetry are separate from your dictation, and a privacy policy should say what they contain.
Then test it. Dictate a harmless, fictional paragraph containing a name, a number and a negation, with Wi-Fi on and then off, and compare the results. Apple's Activity Monitor shows data sent and received while you dictate. A security team can go further with a network inspection tool. Where does my dictation go? maps the paths for SuprFlow specifically.
Questions people ask
Is AI dictation the same as speech recognition?
No. Speech recognition is step 1, turning audio into words. "AI dictation" usually adds a language model that edits those words, and the two steps can run in different places.
Does AI dictation send my voice to the cloud?
It depends on the app and the setting. Some apps recognise speech on the device and send only text for cleanup; others send audio. Check where each step runs before you dictate anything sensitive.
Can AI dictation change what I said?
Yes. The cleanup step is designed to change your wording, and it can change meaning by mistake, especially names, numbers and negations. Read the result before you send it.
Can I turn off the AI part?
Often, yes. Several apps offer a mode with no language-model step, or let you choose a local model for it. Test the setting with a sentence you know, rather than trusting the label.
Related guides
- How speech recognition works
- On-device speech recognition explained
- Local vs cloud formatting
- Does cloud formatting send audio or text?
- What if an AI formatter changes an audit exception?
- Where does my dictation go?
Sources
- Superwhisper Docs — Sensitive Data Best Practices
- Apple Newsroom — Siri AI, a profoundly more capable and personal assistant, is here
- Apple Support — Dictate messages and documents on Mac
- Apple Support — View network activity in Activity Monitor on Mac
- Wispr Flow Help Center — What to expect from Flow: accuracy and known limitations
- Wispr Flow Help Center — Starting your first dictation
- Monologue Docs — Languages and processing models
- Monologue Docs — Permissions and privacy
- BetterDictation — User Guide
- BetterDictation — Privacy Policy