Naming your language instead of leaving it on Auto can improve accuracy, because the model stops spending effort guessing. It helps most on short recordings.
Both rows only appear when your current model supports them. If you cannot
find Translate to English, the model you are on cannot translate. The whole
group is hidden for models with nothing to configure.
Which model for which languages
Parakeet is English only. Nemotron covers 28 languages with automatic detection. Whisper Medium is the widest at 99.
Which models translate to English
Translation is a separate task from recognition. Only some models support it.
Whisper Turbo catches people out. It is a Whisper model, but it does not support translation.
Leave translation off and you get a transcript in the language you spoke. Turn it on and you get English.
Cloud translation
Cloud providers that use the OpenAI multipart format (OpenAI, Groq, and any OpenAI-compatible server you set up yourself) support translation to English through a separate/audio/translations endpoint. Other cloud providers do not.
If translation is on but the selected cloud provider does not support it, SpeakoFlow transcribes normally and logs a warning. The setting row in the cloud section says so in place of a switch that would do nothing.
Simplified and Traditional Chinese
The language list includes both Simplified Chinese and Traditional Chinese. Both are sent to the model aszh, so the model does not distinguish between them.
The distinction matters after transcription. When you select Simplified Chinese, SpeakoFlow converts the output from Traditional to Simplified using OpenCC. When you select Traditional Chinese, it converts the other way. This happens locally, after the model has finished.
Your language choice is only a hint
Your language choice is only a hint
The language you pick is passed to the model only if that model claims to support it. If it does not, SpeakoFlow falls back to autodetection rather than forcing a language the model cannot handle.Picking Japanese while an English-only model is active changes nothing.
Language as a hint for cloud providers
Language as a hint for cloud providers
The language setting is forwarded to cloud providers as a recognition hint:
language_code on ElevenLabs, language on the OpenAI schema and Deepgram.
The transcript still comes back in the language spoken. Script subtags like
zh-Hans and zh-Hant are stripped on the way out, because no cloud endpoint
here understands them.Custom words work differently per model family
Custom words work differently per model family
Whisper-family models receive your custom words as a recognition prompt, which biases the model before it decides. Every other family gets a fuzzy correction pass over the finished transcript instead.Both work. The Whisper route tends to be more accurate because nothing has to be guessed after the fact. See Custom words.
Transcription models
The full catalog and how to choose.
Cloud transcription
Using a cloud service instead of a local model.