Skip to content
Download

Languages and speech recognition models

Choose a dictation language, compare supported speech models, and follow multilingual recording guidance.

Choose a dictation language

Open Home, select Dictation Language, and choose the language you plan to speak. Whisperstream uses Automatic selection by default: it chooses the recommended speech-recognition model and, if needed, asks before downloading it. Wait for the download to finish before dictating.

The model currently in use appears under Voice Recognition Model. Automatic selection uses:

  • NVIDIA Parakeet v2 for English.
  • NVIDIA Parakeet v3 for supported non-English languages.
  • Qwen3 ASR when Parakeet does not support the selected language.

If your current model also supports a new language, Whisperstream may ask whether to keep it or switch to the recommendation. Keeping it turns on Manual selection. If a manually selected model does not support the new language, Whisperstream returns to Automatic and offers the compatible recommendation.

Important: Keep each recording to one language. Parakeet and Qwen3 ASR normally detect the language for each recording. Cohere Transcribe always uses your selected Dictation Language. Changing languages mid-recording can reduce accuracy.

Compare speech-recognition models

ModelUsed forLanguagesApproximate download
NVIDIA Parakeet v2Default English recognitionEnglish661 MB
NVIDIA Parakeet v3Default recognition for supported non-English languages25 languages670 MB
Qwen3 ASRRecognition for languages outside Parakeet's range30 languages987 MB
Cohere TranscribeOptional Manual selection14 languages2.9 GB

After the selected model is downloaded, speech recognition runs on your PC without an internet connection. Selecting a language or model that needs another download opens a confirmation first.

These models produce the raw transcript. For optional cleanup after recognition, see Enhancement providers.

Choose a model manually

Automatic is recommended for most users. To choose a compatible model yourself:

  1. On Home, open Advanced.
  2. Turn on Choose speech model manually. This keeps the compatible model already in use.
  3. Under Voice Recognition Model, choose a model.

The picker appears when more than one model supports the selected language. Cohere Transcribe is available only through Manual selection and requires more memory than the other choices. Choosing Parakeet automatically uses its English or multilingual version as appropriate. Turn off Choose speech model manually to return to the recommended model for your language.

Whisperstream keeps downloaded models when you switch, so switching back does not require another download. To remove one, open the model picker and use Delete beside a downloaded model that is not in use. To remove every unused speech-recognition model, open Advanced and select Delete unused models under Model storage location. A deleted model can be downloaded again later.

Supported languages

Whisperstream offers 39 dictation languages. Default model shows what Automatic selection recommends. Where multiple models are listed, Manual selection lets you choose among them. Cohere Transcribe is never selected automatically.

LanguageSupported modelsDefault model
ArabicQwen3 ASR, Cohere TranscribeQwen3 ASR
BulgarianParakeet v3Parakeet v3
CantoneseQwen3 ASRQwen3 ASR
ChineseQwen3 ASR, Cohere TranscribeQwen3 ASR
CroatianParakeet v3Parakeet v3
CzechParakeet v3, Qwen3 ASRParakeet v3
DanishParakeet v3, Qwen3 ASRParakeet v3
DutchParakeet v3, Qwen3 ASR, Cohere TranscribeParakeet v3
EnglishParakeet v2, Parakeet v3, Qwen3 ASR, Cohere TranscribeParakeet v2
EstonianParakeet v3Parakeet v3
FilipinoQwen3 ASRQwen3 ASR
FinnishParakeet v3, Qwen3 ASRParakeet v3
FrenchParakeet v3, Qwen3 ASR, Cohere TranscribeParakeet v3
GermanParakeet v3, Qwen3 ASR, Cohere TranscribeParakeet v3
GreekParakeet v3, Qwen3 ASR, Cohere TranscribeParakeet v3
HindiQwen3 ASRQwen3 ASR
HungarianParakeet v3, Qwen3 ASRParakeet v3
IndonesianQwen3 ASRQwen3 ASR
ItalianParakeet v3, Qwen3 ASR, Cohere TranscribeParakeet v3
JapaneseQwen3 ASR, Cohere TranscribeQwen3 ASR
KoreanQwen3 ASR, Cohere TranscribeQwen3 ASR
LatvianParakeet v3Parakeet v3
LithuanianParakeet v3Parakeet v3
MacedonianQwen3 ASRQwen3 ASR
MalayQwen3 ASRQwen3 ASR
MalteseParakeet v3Parakeet v3
PersianQwen3 ASRQwen3 ASR
PolishParakeet v3, Qwen3 ASR, Cohere TranscribeParakeet v3
PortugueseParakeet v3, Qwen3 ASR, Cohere TranscribeParakeet v3
RomanianParakeet v3, Qwen3 ASRParakeet v3
RussianParakeet v3, Qwen3 ASRParakeet v3
SlovakParakeet v3Parakeet v3
SlovenianParakeet v3Parakeet v3
SpanishParakeet v3, Qwen3 ASR, Cohere TranscribeParakeet v3
SwedishParakeet v3, Qwen3 ASRParakeet v3
ThaiQwen3 ASRQwen3 ASR
TurkishQwen3 ASRQwen3 ASR
UkrainianParakeet v3Parakeet v3
VietnameseQwen3 ASR, Cohere TranscribeQwen3 ASR

If recognition chooses the wrong language

  1. Confirm that Dictation Language matches the language you are speaking.
  2. Use one language for the entire recording.
  3. If Qwen3 ASR repeatedly detects the wrong language, turn on Choose speech model manually, select Qwen3 ASR, then turn on Force selected language.

Force selected language appears only when Qwen3 ASR is manually selected. It tells Qwen which language to listen for instead of detecting it. Returning to Automatic or choosing another model clears this setting. Cohere Transcribe already uses your selected language and does not show this control. If its output is in the wrong language, check Dictation Language before recording again.