HoldToType

HoldToType turns speech into text without sending anything to the cloud. Hold a keyboard shortcut, say a sentence, let go — the recognised text is inserted where the cursor already was, in any program: an editor, a browser field, a terminal, a chat window. Your clipboard is restored afterwards.

Recognition runs on your own computer. The speech engines — whisper.cpp and sherpa-onnx — ship with the app, and models are downloaded once and stay on disk. While a dictation is happening, no request leaves the machine. There is no account, no subscription and no telemetry.

Each language can be given its own model. Whisper covers about a hundred languages out of the box; narrow models can be twenty times faster and more accurate for a single language, and a streaming model shows the words on screen while you are still speaking. The app can also translate as you dictate: speak one language, type another.

After recognition the text passes through rules you set: replacements for names and in-house terms, a dictionary of rare words, and voice commands such as "new paragraph". Optionally a local language model, run by llama.cpp, cleans up filler words, fixes punctuation or changes the tone — again on your own disk, unless you deliberately point it at an external endpoint.

The program is portable: the folder is the program, models included, and deleting the folder is the whole uninstall. The interface comes in eight languages and five designs, and every setting applies the moment you change it.