Choose a transcription language
Choose the spoken language to guide recognition, or let a supported model detect it. This setting requests transcription in the source language, not translation. Voice and audio files keep independent language choices when you change models or connections.
- Open Voice transcription or Audio file, then its Settings cog.
- Choose Transcription → Spoken language. Search by language name or code.
- Select a supported value and Save. Review both tasks if they share a managed runtime whose model you changed.
Changing a managed model restores its recognition controls while keeping the task language. If that language is unsupported, choose a supported value before transcribing. For Parakeet, use Automatic detection.
Choose a mode
Section titled “Choose a mode”| Choice | What it requests |
|---|---|
| Server default | The server’s configured language or detection behavior |
| Automatic detection | Language detection, where the selected model supports it |
| A named language | Its code, such as en, es, or ja |
| Custom server value | An explicit value your compatible server requires; existing unlisted values stay here |
Specialized model profiles narrow the available choices:
| Model profile | Language choice |
|---|---|
| Nemotron 3.5 ASR streaming | Automatic or a listed locale such as en-US, in completed and realtime mode |
| Qwen3-ASR | Listed hints for completed recordings/files; automatic in realtime. Cantonese/Filipino explicit hints are unavailable with supported vLLM |
| Cohere Transcribe | Select the spoken language for non-English audio; Server default (English) is not automatic detection |
| Parakeet TDT v3 / Voxtral Mini Realtime | Automatic detection |
Custom values, where available, allow at most 32 UTF-8 bytes, without control
characters. Ordinary values are sent unchanged; reserved auto uses Automatic
detection behavior.
Provider behavior
Section titled “Provider behavior”With the Generic model profile, these backend mappings apply:
| Backend | Server default | Automatic detection |
|---|---|---|
| Generic OpenAI-compatible | Server default | Relies on server detection |
| Speaches | Server default | Detects with a supported model, such as Whisper |
| whisper.cpp | Server-configured language | Overrides the server default with detection |
| vLLM | Server default | Detects if the model supports it |
A named language requests that language on all four. Generic, Speaches, and vLLM send the same request for Server default and Automatic detection. Choose Automatic detection with whisper.cpp to override a fixed server language. If detection chooses incorrectly, select the spoken language explicitly where supported.
S1-mini cleanup is English only
Section titled “S1-mini cleanup is English only”S1-mini’s cleanup language cannot be changed. Its handling depends on what is known about the transcript:
| Language information | Cleanup result |
|---|---|
| Explicit non-English selection | Skipped; raw transcript retained |
| Server reports non-English or mixed-language text, even after English was selected | Skipped; raw transcript retained |
| Server default / automatic, with no known language | Runs assuming English, as shown in the controls |
A language mismatch makes no cleanup request. The English-only notice explains why raw text is used for normal safe insertion or explicit file-result copying. Enabled history records the reason within its retention limits.
For non-English cleanup, choose a suitable model with the Generic cleanup profile and instructions, or turn cleanup off. Selecting S1-mini does not change the saved transcription language. Freehand neither infers languages from model names, probes models for language support, nor silently translates text.