Voice Input into Any Windows Window in a Second: Offline, on CPU, Without a Single Gigabyte of Torch

The author developed WhisperType, a free and open-source (MIT) dictation tool for Windows 10/11 x64. It allows voice input into any window via hotkey, works offline without admin rights, and uses a local model deepdml/faster-whisper-large-v3-turbo-ct2 (CTranslate2, int8_float32). The main problem is that the Whisper encoder always processes a fixed 30-second window, so even a short phrase takes about 4 seconds on CPU. The solution is to monkeypatch the pad_or_trim function to narrow the encoder window according to the real speech length. Also, timestamps are disabled when the window is narrowed, and a text-based method for removing rare repetitions is used. For long dictations, stream recognition is implemented: audio is cut at pauses and processed in chunks while recording continues. The optimal number of threads is auto-selected as all logical processors minus two. Additional optimizations: explicit language detection, always-open audio stream, pre-warmed model, deferred imports, and checks for silence and audio device failures. The system-level code uses only ctypes and WinAPI, avoiding hooks libraries.