superwhisper adds words, repeats text or answers a question: find the stage, then fix it
By Ryleigh Newman · Published
When superwhisper produces words you never said, the fault is in one of two places: the voice model that transcribed you, or the AI step that processed the text. Open the dictation in History and compare the raw transcription with the AI result. Whichever one first shows the problem is the stage to fix.
I make YapToText, a different dictation app for the Mac, so read what I say about superwhisper with that in mind. Every fact about superwhisper here is from its own docs, read on October 10, 2026. The page is Gibberish, repeated, or made-up text.
This app is completely free. You don't have to make an account. There are no subscriptions, no difficult downloads, and nothing is locked. Made for Apple silicon Macs. Source on GitHub.
This guide in other languages: العربية, Čeština, Dansk, Deutsch, Ελληνικά, Español, Suomi, Français, עברית, हिन्दी, Magyar, Bahasa Indonesia, Italiano, 日本語, 한국어, Norsk, Nederlands, Polski, Português, Română, Русский, Slovenčina, Svenska, ไทย, Türkçe, Українська, Tiếng Việt, 简体中文, 繁體中文
Which stage did it
From superwhisper's docs, read October 10, 2026.
| What you see | Stage | First thing to try |
|---|---|---|
| Made-up or repeated words in the raw transcription | Voice transcription | Turn on Silence removal |
| The raw transcription is right, the result is wrong | AI processing | Check the mode's prompt, model and context |
| It answered a question, or added comments | AI processing | Tighten the custom instructions |
| The text is in the wrong language | Voice transcription | Set a fixed language |
Fix made-up words in the transcription
- Open Settings, Sound.
- Make sure Silence removal is on. It cuts silent stretches before transcription, and is on by default in macOS version 2.18.3 and later.
- If it continues, open Vocabulary and keep only the terms you need.
The docs name two causes at this stage. Silence can make a transcription model generate words that were never spoken, and so can a vocabulary list with too many entries. The page on Whisper hallucinations explains why speech models do this at all.
Fix a wrong AI result
- The prompt. In a custom mode the instructions are yours. Imprecise instructions, no examples, or directions that conflict are the usual causes.
- The model. The docs say larger cloud models handle complex requests with fewer hallucinations, and smaller local models need more specific prompting. Try a more capable model for that mode.
- The context. A text field full of unrelated text, or something you copied by accident while dictating, gets sent as context and can mislead the AI. Turn on context only in modes that need it.
The modes and prompts page covers writing instructions and what each context type reads.
If you only want your words
The Voice to Text mode has no AI processing, so the second stage cannot change anything. That leaves only the transcription stage to get right.
The same problem in YapToText
YapToText runs Whisper too, so it can meet the same first-stage problem: a long silence can produce a stray phrase. Raw Transcription mode skips the cleanup model entirely and gives you exactly what the speech model heard, and History keeps the raw transcript next to the cleaned one for every dictation, so you can see which stage changed what. The page on dictation without AI covers the no-rewrite setup.
This app is completely free. You don't have to make an account. There are no subscriptions, no difficult downloads, and nothing is locked. Made for Apple silicon Macs. Source on GitHub.
The short version
- Compare. Raw transcription against AI result, in History.
- Transcription stage. Silence removal on, vocabulary short.
- AI stage. Clearer prompt, stronger model, less context.
- No AI. Voice to Text mode.
- Wrong language. A fixed language in the mode.
If superwhisper changes these settings, open an issue on GitHub and I will update this page. Every one gets read.
This app is completely free. You don't have to make an account. There are no subscriptions, no difficult downloads, and nothing is locked. Made for Apple silicon Macs. Source on GitHub.
Related
Sources
- https://superwhisper.com/docs/common-issues/hallucinations (checked 2026-10-10)
- https://superwhisper.com/docs/common-issues/context (checked 2026-10-10)
- https://superwhisper.com/docs/get-started/history (checked 2026-10-10)
- https://github.com/ryleighnewman/YapToText (checked 2026-10-10)
- https://apps.apple.com/us/app/yaptotext/id6786382289 (checked 2026-10-10)
Questions
Why does superwhisper add words I did not say?
Its docs name silence and an overlong vocabulary list as the causes at the transcription stage. Turn on Silence removal and trim the list.
Why does superwhisper repeat the same phrase?
That is a transcription hallucination. Check Silence removal in Settings, Sound, and shorten your vocabulary list.
Why did superwhisper answer my question?
The AI step treated your dictation as a request. Tighten the mode's instructions, or use a mode without AI processing.
How do I see what superwhisper heard?
Open the dictation in History. It shows the raw transcription and the AI result.
Where is Silence removal in superwhisper?
In Settings, Sound. It is on by default in macOS version 2.18.3 and later.