Dictation for developers on Mac
Developers write more prose than they expect: prompts for AI assistants, commit messages, pull request descriptions, comments and docs. WhisprDesk lets you speak that text into any text field where you can paste, using one hotkey, so your hands stay on the work.
Long prompts are easier to say than to type
A good prompt for an AI assistant often runs to several sentences of context. Press the hotkey, explain the problem the way you would to a colleague, and the text lands in the chat box where your cursor is. You can then edit it like any other text.
The same applies to issue descriptions, design notes and the explanation at the top of a pull request. If you can paste into the field, you can dictate into it.
Commit messages, comments and docs
Click into the text field, press the hotkey, say what changed and why, and press it again to stop. Dictation produces ordinary prose, so you will usually want to trim it into your team's commit style before you save. Code itself is better typed; dictation is for the words around it.
Add project names, library names and internal terms on the Vocabulary tab so they are recognized more reliably. Vocabulary helps Whisper and cloud models; NVIDIA Parakeet does not use it. With cloud transcription, your Vocabulary words are sent to your provider along with your audio.
Where the text lands
WhisprDesk pastes into whichever app has focus, so it works in any text field where you can paste. It uses the Accessibility permission to do that. If pasting is blocked, for example in a secure field such as a password box, the text stays on your clipboard so you can paste it yourself with Cmd+V. After pasting, your previous clipboard contents are restored unless you turn on Keep transcription in clipboard in Settings > Preferences.
Give an instruction instead of dictating
With AI cleanup turned on (Models > AI cleanup > Enable text cleanup, plus a cloud key or a local model), you can start with "Hey" and the voice command name (Models > AI cleanup > Voice commands; the default is WhisprDesk) and say what you want, for example "Hey WhisprDesk, make this sound more formal." The AI follows the instruction and pastes the result. Without it, cleanup only tidies your words. Cleanup is optional and off by default.
You can change the name under Voice commands in Models > AI cleanup. Saving a new name also adds it to your Vocabulary. Prompt Studio, further down the same page, lets you view, customize and test the instructions cleanup uses, so you can shape how it handles technical text. Testing a prompt uses your cleanup model the same way a dictation does.
Agent Mode is separate. It is an optional voice chat window with its own model and hotkey, set up in Models > Agent, and it is off by default.
A local API for your own scripts
WhisprDesk includes a local API that only accepts connections from your own Mac and is protected by a token. You find the endpoint and token in Integrations > External App Gateway, where you can also rotate the token. Copy the endpoint from the app rather than typing it, because it can change if another app is using the usual one.
The API has five endpoints. Status and capabilities need no token and confirm the API is running and list the local engines and limits. Transcribe takes an audio file and returns the text, the engine and model used, and the duration. History returns your recent saved transcripts, and Events is a live stream that fires each time a dictation is saved to Your Transcripts.
To transcribe, send the raw audio bytes as the request body with an audio Content-Type, not a multipart form or JSON. Files can be up to 25 MB. Transcription through the API always runs on a local engine. Whisper requests always use the Base model, so download Base first or the request fails, and Parakeet requests use Parakeet TDT 0.6B. Transcribe needs an active license or trial.
Requests are limited to 60 per minute and 10,000 per day. Errors include 401 for a missing or wrong token, 402 for no active license or trial, 413 for a file over 25 MB, 415 for an unsupported Content-Type, 422 when no speech is detected, 429 for too many requests, and 500 or 503 when transcription fails or the engine is not available. A request can also be rejected for an empty body, a forbidden host or an unknown path. API transcriptions are not saved to Your Transcripts, and with Data Retention off, History and Events receive nothing new. The Local API page has the full reference and a worked example.
Privacy for work you cannot share
Where your voice goes depends on the engine you choose. With a local model (NVIDIA Parakeet is about 680 MB), transcription runs on your Mac and your audio is not uploaded. With a cloud provider, your audio goes directly to the provider whose key you entered. WhisprDesk does not receive it either way.
If your employer restricts where source code or internal discussion can go, choose a local model and keep AI cleanup off, or use a local cleanup model. With cloud transcription, your audio and your Vocabulary words, such as internal project names, go to your provider. With cloud cleanup, the transcribed text goes to that provider, not your audio. Check your own policy; WhisprDesk cannot make that call for you.
Frequently asked questions
- Does it work inside my editor or terminal?
- WhisprDesk pastes text into whichever app has focus, so it works in any text field where you can paste. It needs the Accessibility permission. Some secure fields, such as password boxes, can block it, and if pasting is blocked the text stays on your clipboard for a manual Cmd+V.
- Can it dictate code?
- It transcribes speech into prose. Names and technical terms can be added on the Vocabulary tab, but typing is still the better tool for code itself.
- Does my code or voice leave my Mac?
- Not with local transcription. With cloud transcription your audio goes to the provider you chose, and cloud AI cleanup sends transcript text to its provider. You pick each one.
- Can my own scripts use WhisprDesk?
- Yes, through the local API. It only accepts connections from your own Mac and needs a token from Integrations > External App Gateway. Transcription through it runs on a local engine.
- Can I change how AI cleanup treats technical text?
- Yes. Prompt Studio in Models > AI cleanup lets you view, customize and test the cleanup instructions.