HoldToType

HoldToType turns speech into text at the cursor in any Windows program: hold two keys, speak, let go, and the words land where you were typing. Recognition runs on your own computer with open models, and you are not tied to Whisper: Nemotron 3.5 shows live text while you speak, Whisper covers 99 languages and translation, Parakeet serves 25 European languages, Canary translates between English, German, Spanish and French, and GigaAM, Qwen3-ASR and Moonshine cover single languages fast. The program keeps one model per language and switches by itself; your own Whisper or sherpa-onnx models can be dropped into the models folder.

Nothing leaves the machine: no account, no subscription, no credits, no telemetry. A small editor model on your disk (any GGUF, run by llama.cpp) can remove fillers, fix punctuation and change the tone, with prompts you write yourself. Replacements, a dictionary of rare words and voice commands teach it your vocabulary; translation into seven languages is one setting away.

It comes as an installer or a portable folder you can carry on a stick, with seven designs and eight interface languages. MIT licence, source on GitHub, builds reproducibly with one script and Docker.