Voice Input

Voice input lets you dictate instead of type. Hold a hotkey, speak, release — the transcript lands wherever you've told it to go. The whole pipeline runs locally: audio capture, resampling, and transcription all happen on your machine, with no network call in between.

If you have ever dictated into a cloud speech API and wondered where that audio ended up, this is the other design. Your voice is captured, transcribed by a local Whisper model, and written into MadoHub — all in the same process, on your machine. Nothing about the audio or the transcript is uploaded.

What you can do with it

• Push-to-talk dictation — hold a hotkey, speak, release to transcribe.
• Retry the last recording — re-run transcription on the same captured audio without speaking again, useful when the first pass misheard you.
• Route the transcript to one of three destinations — into the MadoAgent chat input, into the focused terminal without submitting, or into the focused terminal and submit immediately.
• Switch destinations on the fly — change where dictation lands from the command palette without opening Settings.
• Pick a model size — from tiny (~75 MB) up to medium (~1.5 GB); larger models are more accurate but slower and heavier to download.
• Pin a language or auto-detect — English, Chinese, Japanese, Korean, Spanish, French, German, or auto.

How to use it

1. Enable voice input from Settings → Voice, or from the command palette (Cmd+Shift+P) via the Enable Voice Input entry (searchable by "voice", "dictation", "push-to-talk", or "语音").
2. If you have not downloaded a Whisper model yet, MadoHub opens Settings on the Voice tab with a toast telling you what to download — it will not record into silence.
3. Hold Cmd+Shift+V (the default push-to-talk hotkey) to record, and release to stop and transcribe.
4. Press Cmd+Shift+R to retry the last recording if the first pass misheard you. Retry only works within 30 seconds of a successful transcription and only if you have not started a new recording since.
5. Choose where transcripts land in Settings → Voice, or switch destinations on the fly from the command palette.

The two terminal destinations lock to whichever terminal tile was focused when you started recording — a focus change mid-recording cannot misroute the transcript. If no terminal was focused (or the tile closed before you finished speaking), MadoHub falls back to the Agent destination and shows a toast explaining why.

Common use cases

• "Refactor the auth module to use a class, add tests for the login flow." — dictate a long prompt into MadoAgent while walking through the code in your head, then review and send.
• Running a quick shell command without typing — set the destination to Terminal — auto and dictate git status or pnpm test.
• CJK dictation — pin the language to Chinese/Japanese/Korean for better accuracy on short utterances, since Whisper's auto-detect can misfire on quiet or very short CJK clips.

Tips and best practices

• Start with the small model (the default) for a good accuracy/size tradeoff. Move up to medium only if you dictate often and need the extra accuracy.
• Pin a language when you know what you are speaking — it avoids auto-detect mistakes on short or quiet clips.
• Rebind the push-to-talk and retry hotkeys in Settings if they collide with shortcuts you already use; Settings warns you about conflicts.
• Use Terminal — insert rather than Terminal — auto when dictating multi-line input, so you can review before submitting.

Related features

• Command Palette — toggle voice input and switch destinations from here.
• Settings — the Voice tab where you configure model, language, hotkeys, and destination.
• MadoAgent — the default dictation destination.

Limitations

• Voice input ships off by default; enable it before first use.
• Retry only works within 30 seconds of a successful transcription, and only if you have not started a new recording since.
• Terminal destinations require a live terminal tile focused at the moment you start recording. Otherwise MadoHub falls back to the Agent destination.
• A model must be downloaded before recording; MadoHub will not record into silence.