Transcription Settings
Settings > Transcription controls what the speech-to-text engine produces, independent of any AI refinement applied afterwards.
Language
A picker with Auto-detect (the default) plus the full list of supported languages — common ones first (English, Spanish, French, German, Italian, Portuguese, Chinese, Japanese, Korean, Hindi, Arabic, Russian), then everything else alphabetically.
- Auto-detect works well for Whisper and OpenAI models and is the right choice if you dictate in one language at a time but switch between languages day to day.
- Forcing a language noticeably improves accuracy when auto-detection guesses wrong — most common with short dictations, heavy accents, or languages that sound similar.
How much choice you actually have depends on the active model:
| Model | Language behavior |
|---|---|
| Whisper multilingual | 99 languages, auto-detect or forced |
Whisper .en variants |
English only |
| Apple Speech | Pick a locale (for example "English (US)") — required |
| Parakeet V2 | English only |
| Parakeet V3 | Auto-detects 25 European languages — no manual choice |
| OpenAI | 99 languages, auto-detect or forced |
See Engine Selection for choosing a model by language needs.
Transcription Prompt
A free-text prompt passed to the speech engine to steer how it transcribes. Only Whisper and OpenAI models support prompts — Apple Speech and Parakeet ignore this setting.
Prompts influence three things well:
- Spelling hints — list unusual names and terms so they come out right:
ZyntriQix, VortiQore V8, WhisperKit - Punctuation style — write a sample sentence with the punctuation you want:
Hello, welcome to my lecture. - Keeping filler words — show fillers in the prompt and the engine preserves them:
Umm, let me think like, hmm...
A Clear Prompt button appears whenever the prompt is not empty.
A prompt is a bias, not a command. The engine treats it as preceding context and tends to imitate its style and vocabulary — it cannot follow arbitrary instructions like "summarize" or "translate." Keep it short; only the trailing portion is used by the engine.
Prompt or custom instructions?
Spelling and punctuation problems are best fixed here, at the transcription layer. Tone, structure, and formatting belong in your workflow's custom instructions, which run after transcription. If a word is heard wrong, prompt it; if it is heard right but written wrong for the context, use a workflow.