Language and prompts

Which language to listen for, and how to steer the words. Mostly Whisper; notes say where else each applies.

Language detection

English-only speakers can use the smaller .en models; for multi-language detection, pick a model without the .en suffix:

{
    "language": null // null = auto-detect (default), or specify language code
}

Auto-detect (null, the default) runs an extra detection pass per utterance. Setting language explicitly skips it and lowers latency.

Whisper accepts any of its 99 supported language codes — the core set:

Code	Language	Code	Language
en	English	pt	Portuguese
de	German	nl	Dutch
fr	French	pl	Polish
es	Spanish	ru	Russian
it	Italian	zh	Chinese
ja	Japanese	ko	Korean

For the full list, see the Whisper language codes.

Whisper prompt

Customize transcription behavior:

{
    "whisper_prompt": "Transcribe as technical documentation, keeping acronyms uppercase."
}

The prompt influences how Whisper interprets and transcribes your audio, eg:

• "Transcribe as technical documentation with proper capitalization, acronyms and technical terminology."
• "Transcribe as casual conversation with natural speech patterns."
• "Transcribe as an ornery pirate on the cusp of scurvy."

A whisper_prompt applies to every language, and a prompt written in one language pulls transcription toward that language. That is why the shipped capitalization default lives in whisper_prompt_en instead — see below.

Translation

Translate non-English speech into English:

{
    "task": "translate",
    "language": "it"  // optional: set source language, or null to auto-detect
}

• "transcribe" (default) - Output in the source language
• "translate" - Translate speech into English

Note: Supported by faster-whisper and pywhispercpp backends. qwen3-asr honours language but has no translation task. language and task are independent — setting a non-English language does not imply translation.

Language-specific prompts

Set a per-language prompt using whisper_prompt_{lang}:

{
    "whisper_prompt_de": "Transkribiere auf Deutsch. Verwende Schweizer Rechtschreibung: kein ß, immer ss."
}

• The language comes from language, secondary_language or --lang; pywhispercpp and faster-whisper auto-detect it when unset
• Prompts do not apply to qwen3-asr, which sends no prompt field (it uses language for the hint and auto-detects when unset)
• Falls back to whisper_prompt if no language-specific prompt is configured
• With neither set, English audio gets the shipped whisper_prompt_en capitalization prompt and other languages get no prompt

Decoding strategy

Controls how Whisper searches for the best transcription. Applies to pywhispercpp and faster-whisper backends.

{
    "sampling_strategy": "beam_search",  // "beam_search" (default) or "greedy"
    "beam_size": 5                       // number of candidates to track (beam_search only)
}

• "beam_search" (default) — keeps the top N candidate sequences in parallel and picks the best overall result. Matches whisper-cli defaults. Better accuracy, especially for non-English audio and noisy input.
• "greedy" — picks the single highest-probability word at each step. Faster, lower quality.
• beam_size — higher values (e.g. 8–10) can improve accuracy at the cost of speed. Default 5 is a good balance for real-time dictation.

Note: sampling_strategy is locked in at model load time for pywhispercpp. Changing it requires a service restart.