Choose a speech engine
- Apple Speech can provide live results through the system speech framework when on-device recognition is available for your language and device.
- WhisperKit offers downloaded Whisper models for multilingual on-device transcription.
- Parakeet provides local recognition for supported languages. It returns the transcript after processing rather than a streaming preview.
- Optional cloud engines, such as Meta with Muse Voice Transcribe, are used only after you configure and select a provider.
Note
Local is the default
With an on-device engine, recording and transcription stay on the device. Choosing a cloud engine sends audio to that selected provider for transcription.
Build a useful profile
Name the purpose
Use a task name such as Quick Note, Meeting Notes, Translate to English, or Focused Writing.
Set language, engine, and model
Choose an explicit model when consistency matters, or use the automatic route when you want TypeWhisper to select a compatible option.
Choose translation and Whisper Mode
Set a translation target when needed. Whisper Mode is intended for quieter speech and changes how capture is handled.
Choose the writing result
Keep the raw transcript, apply a formatting instruction, or use an available structured writing action.
Review before you reuse
TypeWhisper keeps the original recognition result alongside processed text. That makes it possible to compare what the speech engine heard with the formatted or translated result.
Apple Translate can translate dictated text on device. Premium writing features can improve a fresh transcript, learn confirmed corrections, and create reviewable email drafts with a subject and body.
Choose a writing service
Formatting instructions and structured writing actions run through a writing service. Choose it under Settings > Writing Services.
- Apple Intelligence processes writing text on the device where it is available.
- Downloaded local models run entirely on the device and do not require Apple Intelligence. TypeWhisper offers LFM2.5 2.6B (about 1.7 GB) and Gemma 4 E2B (about 2.8 GB). You review the model license before the download starts.
- Cloud services such as OpenAI, OpenRouter, and Meta with Muse Spark run under your own account with that provider. The Meta API key is shared with Meta speech recognition.
- Swipe the row of a local model to pause or remove its download.
Note
What a writing service receives
Local models and Apple Intelligence keep writing text on the device. A cloud service receives only the text needed for the selected writing task, not audio, History, or the dictionary.