The short answer
Yes, VoiceKeyboard recognises speech without the internet. Switch RECOGNITION to LOCAL, pick a model, and the text is typed into any window from your microphone with the network cable pulled out. The FILES screen, which turns recordings into text, always works this way. The internet is needed once, to download the model.
Which model for your language
There are five offline models. Each one is good at something different:
| Model | Download | Text appears | Languages |
|---|---|---|---|
| Whisper | 75 MB to 1,031 MB (five sizes) | after each phrase | 58 to choose; about 99 on AUTO, which detects the language |
| Parakeet v3 | about 670 MB | after each phrase | 24 European languages, detects which one |
| Vosk | about 40–50 MB per language | while you speak | 15, one small model per language |
| GigaAM v3 | about 232 MB | after each phrase | Russian, with punctuation |
| T-one | about 144 MB | while you speak | Russian, without punctuation |
- English and Spanish: Whisper, Parakeet v3 and Vosk.
- Russian: all five; GigaAM v3 and T-one are made for it.
- Japanese and Hindi: Whisper and Vosk.
- German, French, Italian, Polish, Ukrainian and others: Whisper for all of them, Parakeet v3 for most European ones, Vosk for some.
Not sure? Leave LOCAL on "Auto — by language", the first entry in the model list. For the language you dictate in, it runs the best model you have downloaded, and if you have none yet, it offers the one that suits your languages.
Your hardware matters for one model
Whisper is the most capable and the most demanding. With an NVIDIA graphics card it runs several times faster through CUDA; on an ordinary processor it is slow, and for live dictation it can lag behind your speech. On a laptop without an NVIDIA card, pick Parakeet v3, GigaAM v3, T-one or Vosk for dictation. They run on the processor without trouble.
Vosk and T-one show words while you are still speaking. Whisper, Parakeet and GigaAM wait for a pause and then type the whole phrase, usually with better punctuation.
Files and recordings offline
FILES transcribes with Whisper on your PC, every time. A recording from disk needs no internet once the model and FFmpeg (a free converter the app offers to download) are there. A video link is different: the sound has to be downloaded first, then recognition runs offline. Details are in transcribing audio files offline.
What still needs the internet
- Downloading a model, once. They come from their official sources, with a mirror on voice-keyboard.com as a fallback.
- Google and Soniox, the two online engines. Switch to LOCAL and they are not used.
- Video links on FILES, for the download.
Pro also works offline, until the end of the period you paid for.
Why offline
With a local model the audio never leaves the computer: the model listens, types and forgets. That matters for client calls, patient notes, legal drafts and code under NDA, and for anyone whose employer does not allow cloud speech services. It also means no lag from a slow connection and no dictation outage on a plane.
The trade-off is honest to state: the online engines are usually stronger on hard audio, and Soniox detects the language as you speak, even mid-sentence. Many people keep both, LOCAL by default and ONLINE when it helps. The online and local engines page compares them.
Offline compared with Windows voice typing
Windows 10 and 11 have voice typing built in (Win+H), but Microsoft's help page says it "uses online speech recognition, which is powered by Azure Speech services". Without a connection, it does not work. The full comparison is in Windows voice typing vs VoiceKeyboard.
Try it
Install VoiceKeyboard, switch RECOGNITION to LOCAL, and download a model for your language. Then turn off Wi-Fi, put the cursor in any window and press Ctrl+1.
