DictatorFlow Update: Custom Wake Words, Button-Free Voice Input

DictatorFlow recently updated its voice input method. Developer Lee Penkman demonstrated on X that you can now train custom wake words and go completely button-free. The app is free, cloud transcription is optional, and you can also bring your own API key.

DictatorFlow is a cross-platform voice dictation app that supports Windows, macOS, and Linux. Native binaries written in Zig, without the memory overhead of Electron. It runs locally offline, and audio never leaves the device. It also offers a cloud transcription mode that automatically switches between multiple providers. No account is needed to download, and local inference is unlimited.

Lee Penkman's demo video:

The wake word feature is accessed within the app. Train your own wake word, then just speak, and the text will be entered into the currently focused window. Previously you needed to press the Ctrl+Shift+D shortcut; now you don't need to press any key.

The official website also has a section called "Measured, Not Marketed". The developer publicly discloses the WER performance of every speed optimization, including experiments that haven't shipped. For example, the WER of the VibeVoice-ASR CPU path is 60.45%, so it is not a default option. This kind of transparency is rare in voice products.

Other notable features include voice editing. Select text in any app and say "Make it more professional", and the result replaces it directly. It can also translate, summarize, and modify code. A browser widget can add voice input next to any textarea. The API supports PCM, WAV, WebM, MP3, and OGG, billed by the second at $0.004/second.

Technically, the local model is Parakeet CTC 0.6B, running on ONNX Runtime, supporting NVIDIA CUDA and AMD ROCm, and it can also run on CPU. The entire project is open source under Apache-2.0, including the desktop app, local engine, and API server. The cloud service is paid, and the revenue supports the free local version.

For people tired of keyboard input or who care about voice privacy, DictatorFlow offers a verifiable option. Open source and data transparency are what make it least like a marketing account.