Transcript post-processing

Turn on Cleanup to edit a transcript before delivery. Voice and audio files share this optional stage. It sends transcript text to your selected chat service, separate from speech recognition; it does not send the original audio.

Choose a workflow

Desired result	Configuration
Leave transcription unchanged	Keep Cleanup off; no chat connection is needed
Edit using your instructions	Choose a compatible chat connection/model and the Generic model profile
Use English cleanup presets	Choose S1-mini by Superwhisper, through managed llama.cpp or a supported manual service

Raw text is the recovery result

If cleanup fails, times out, is unavailable, or returns empty text, Freehand falls back to the raw transcript. Normal destination checks and manual-copy settings still apply; cancellation prevents automatic insertion.

Compare before relying on cleanup

Cleanup can make mistakes. Enable History before a run to compare raw and cleaned versions after success. History is off by default, limited to 20 entries and 2 MiB in memory, and cleared on exit. Raw fallback on failure does not require history.

Custom-instruction processor

1. Open Voice transcription or Audio file → Settings → Cleanup. Select a saved cleanup connection, or Add connection… using the server connection guide.
2. Enable cleanup and choose a model. Refresh models reads metadata; it does not test cleanup quality. You can also enter the exact model ID.
3. Choose Generic under Model profile. Write the Custom system instruction, or restore the recommended instruction as a starting point.
4. Save, then try a short transcription. Review the cleaned result before relying on it for longer text.

Write how to edit the text without a transcript placeholder: Freehand sends the raw transcript separately with the selected connection’s credentials. Instructions are saved locally; API keys stay in Windows Credential Manager or macOS Keychain.

For manual connections, choose the model profile explicitly. Freehand does not identify S1-mini from its name. Switching profiles preserves custom instructions and S1-mini choices for when you switch back.

S1-mini by Superwhisper with llama.cpp

Managed on this computer

Follow local llama.cpp setup, then select its built-in connection in Cleanup. The runtime guide lists supported platforms. Local transcription and cleanup use separate runtimes and can run together.

Manual service

Select S1-mini by Superwhisper as the model profile. Use a qualified llama.cpp or vLLM deployment; for an existing Generic endpoint, disable reasoning on the server. The S1-mini guide links launch recipes and explains style, structure, and context controls.

S1-mini requires English and reasoning off. Review the language behavior below before using automatic language detection.

Generation controls

Open Cleanup → Generation controls and save before the next recording or file job. Active work retains its settings even if cleanup has not started yet.

Control	Values and effect
Limit output tokens	Optional for Generic, llama.cpp, and vLLM; starts at 2,048 and accepts 1–65,536 whole tokens, not words
Output limit off	Uses the server’s limit; keeps a valid value for later
Disable reasoning with Generic model profile	Optional with llama.cpp/vLLM; asks for a result without a thinking phase. Off uses server behavior
S1-mini Required off	Freehand requests reasoning off with llama.cpp/vLLM
S1-mini Disable on server	With a Generic backend, configure reasoning off yourself

An output limit neither enlarges the context window nor splits long inputs; request timeout is independent. Allow room for the complete cleaned transcript. Support for optional reasoning control depends on the server and model template. Turn a saved custom reasoning override off before switching the backend from llama.cpp/vLLM to Generic.

Review long or limited output

A provider rejection or reported truncation falls back to raw text without retrying other controls. Freehand cannot detect omissions a server does not report. Compare against the raw version when history is enabled.

See llama.cpp generation controls for qualified versions and model limits.

vLLM cleanup

Choose vLLM as the connection’s backend profile for a supported v0.28.0 server. Configure its /v1 endpoint and text model separately from transcription. The same generation controls apply, including required reasoning off for S1-mini. See the vLLM guide for deployment options.

Language compatibility

S1-mini input language	Behavior
English	Eligible for cleanup
Explicit or server-reported non-English / mixed language	Skips cleanup and keeps raw text
Unknown	Runs assuming English, as displayed in its controls

Custom cleanup language support depends on your model and instruction. Language selection explains the complete S1-mini rules.

Choose model controlsUnderstand S1-mini styles, structures, contexts, and supported hosts.

Compare transcript versionsKeep raw and cleaned results for the current session.