Local speech to text for Windows, with AI clean-up on your PC
Short answer: VoiceWink turns your voice into text on your PC, and its built-in AI can clean the text up there too, with no account. By default, only the essentials go online.
Speech to text on your PC
- NVIDIA Parakeet: 25 European languages.
- Whisper: about 100 languages.
Both detect which language you speak, and both use your graphics card when they can (through Vulkan), otherwise the processor. Filler sounds such as “uh” are removed.
AI clean-up on your PC
AI enhancement fixes grammar, drops stutters and false starts, and can rewrite the text for the app you are in: an email in Outlook, a chat message, a prompt for a coding agent. It runs on your PC with a built-in model you download once, in the app. No API key or account.
| Built-in AI model | Download | Needs |
|---|---|---|
| Qwen3.5 2B | 1.3 GB | 16 GB of memory |
| Qwen3.5 4B | 2.7 GB | 16 GB of memory |
| Qwen3.5 9B | 5.7 GB | 16 GB of memory and an 8 GB graphics card |
| Gemma 4 12B | 6.7 GB | 16 GB of memory and a 10 GB graphics card |
Setup offers the model that suits your PC, and the app shows each model’s speed on your PC. You can also use your own AI server (Ollama, LM Studio) or a cloud provider.
What still goes online
By default, only the essentials go online, such as the license check, updates and model downloads. The privacy policy lists everything.
VoiceWink is open source (GPL v3).
Without internet
Dictation and AI enhancement work without internet once the models are downloaded. A paid license checks in online at least once every 30 days. Cloud features need internet.
What your PC needs
- Windows 11 (x64 or arm64).
- Built-in AI: 16 GB of memory; the two largest models also need a graphics card (see the table).
- Downloads: 44 MB to 0.9 GB per speech model, 1.3 to 6.7 GB per AI model.
When the cloud is the better choice
- Your PC has less than 16 GB of memory. Speech to text still runs on your PC; use cloud AI enhancement or your own AI server.
- The built-in models are too slow on your PC. A cloud model can answer faster.
Cloud features are optional and use your own API key, at the provider’s prices. Your audio or text then goes to that provider.
- Transcription: Groq, Deepgram, ElevenLabs or OpenAI.
- AI enhancement: Anthropic, OpenAI, Gemini, Groq, Mistral, OpenRouter or Cerebras.
- Voice-to-image: OpenAI, Gemini or OpenRouter.
Other ways to dictate without the cloud on Windows
- Windows voice access, in Windows 11 (22H2 and later): “on-device speech recognition” that “works even without the internet”, after a setup that downloads language files. Seven languages, some only in certain regions. It also controls the PC by voice.
- Fluid dictation, on Copilot+ PCs, in English: it “automatically corrects grammar, punctuation, and filler words as you speak”, with “on-device small language models (SLMs)”. In voice access, and in voice typing, whose speech recognition is online.
More: VoiceWink vs Windows voice typing, dictation apps for Windows, compared.
Price and trial
$25 once for one PC, VAT included. No subscription; updates are free when released. Free for 7 days: no license key, email or credit card. 30-day money-back guarantee.
Sources
Written by DV Studio, the maker of VoiceWink, on 12 October 2026, from the app’s code and its privacy policy. Microsoft’s facts come from its own Support pages, read that day: voice access, voice access setup, Fluid dictation, voice typing. Features may have changed since. Product names are trademarks of their owners; VoiceWink is not affiliated with, endorsed by or sponsored by any of them.