Handy puts local speech transcription into the text field you are already using
CJ Pais built an accessibility tool people can extend themselves. Optional AI cleanup introduces a separate privacy choice after speech recognition finishes.

Your speech can become text without making a round trip through a cloud transcription service. Handy lets you hold a shortcut while speaking or tap it to start and stop, then sends the transcript to the text field you already have open. The editor doesn’t need to know it just became a dictation app.
Developer CJ Pais (@cj_pais) states the motivation in the project’s README: “Accessibility tooling belongs in everyone’s hands, not behind a paywall.” The same introduction stresses extensibility. A useful speech tool can also be a starting point for somebody else’s version, rather than another service whose behavior is fixed behind an account screen.
The project has a predecessor. Pais’s original Python command-line Handy combined local MLX Whisper transcription with keyboard automation, voice-triggered code generation and AI assistance through OpenRouter. Its different shortcuts selected different jobs. The desktop project’s basic promise is easier to explain: speak, then put the words where the cursor is waiting.
Underneath that shortcut, Silero voice-activity detection filters silence before transcription, according to the README. Whisper models offer GPU acceleration where available; Parakeet V3 provides a CPU-optimized option. These are choices about what your computer runs, so an offline label doesn’t tell you which model will feel fast enough on the machine under your desk.
The FAQ adds two limits that matter more than a long model menu. Handy currently takes microphone input rather than system audio, and, as the FAQ puts it, “Handy records a segment, then transcribes it.” It isn’t a live-caption overlay for everything your computer plays. It also recommends keeping recordings under three to five minutes. A quick dictated reply and an hour-long meeting are different requests.
Getting the transcript into the right app can be the fiddly part. Handy’s paste-method documentation describes the default path as copying text to the clipboard and simulating Cmd-V or Ctrl-V. Windows Terminal, Command Prompt and PowerShell get a different recommendation: Shift-Insert. When a sentence fails to arrive, the receiving application’s paste behavior deserves attention before you blame speech recognition.
The same page offers Direct Input, typing characters without the clipboard, but warns that non-US layouts can produce garbled output. Its None option copies the result without pasting, useful when you want to change windows first. Recognizing every word correctly won’t help if the transcript lands in the wrong conversation.
The privacy decision continues after recognition. The FAQ says normal network requests download models and check for updates, but enabling a cloud provider for post-processing sends transcriptions to that provider’s API. A local microphone-to-text pipeline can therefore feed a remote cleanup step if you choose one.
The post-processing guide describes an experimental option for grammar fixes, translation and custom transformations. It can use an on-device provider or a custom local endpoint as well as cloud services. Crucially, this feature has a dedicated shortcut; the normal transcription shortcut returns plain output. That separation lets a user decide whether this particular sentence needs another model involved.
Tools that separate recognition, cleanup and delivery could give users more useful control over dictation. They also need to keep those choices understandable when the application is hidden in a tray and a shortcut is doing the work. The next practical test is a paragraph full of your own names and technical terms, followed by the question no model list answers: how much correcting remains?