Alpha for macOS and Linux · runs on your computer

Vim for your voice.

Hold Talk on your phone and speak. Vispr types what you said wherever your cursor is and tidies it up as you go. Then use the same phone to move, select, fix and send, in any app or coding agent. No controller needed, though a DualSense works great too.

Take the tour
agent · ~/projects/vispr AGENT

╭─ deepseek-v4.1-flash · session 3 ─────────

● Updated live.rs: the tick loop no longer awaits the insert task.

└ 4 files changed, 38 tests passed

›

⏵⏵ accept edits on · shift+tab to cycle

Talk Hold Talk on your phone and speak
~70msfrom release to the first formatted word
0bytesof your voice leave the computer
120Hzinput loop, never blocked by the models
0controllersneeded: your phone is the remote

Formats as you watch

Speak naturally.
Get clean text.

Vispr transcribes while you're still talking. When you let go, a small model made for dictation tidies the transcript: it adds punctuation, drops filler words, applies corrections like “Tuesday, I mean Wednesday”, and formats lists. The first formatted word appears about 70 ms after you let go.

  • RawRaw mode. Tap Raw on your phone mid-sentence to keep your exact words (△ on a DualSense). Vispr remembers.
  • UndoUndo. The button beside Talk removes exactly what you just inserted (○ on a DualSense).

TEXT Listening Raw

Features

Dictation that fits your work.

Phone

Your phone is the remote

Scan a code once and your phone becomes a trackpad, a slide pad for the caret, the keys each mode needs and a big Talk button. It connects straight to your computer over your Wi-Fi. Nothing to buy.

Anywhere you type

Types where your cursor is

Editors, browsers, chat apps and terminals. On Hyprland and macOS it cancels instead if focus moves first.

Modes

A mode for each job

Tap the mode chip to switch between editing text, working in a terminal agent and moving around Herdr.

Editing

Fix it by voice

Slide to move the caret, hold and slide to select, then talk over the selection to replace it.

Controller

Got a DualSense? Even better

Sticks under your thumbs, its own microphone and a button for every action. Use it on its own or alongside your phone.

Private

Runs on your computer

Speech recognition and formatting both run locally, on open models. No account, no cloud, no telemetry. Models download once and are checked against pinned checksums. Your phone talks to your computer directly over your Wi-Fi.

The tour

Four things you'll do all day.

All on your phone. Scroll to step through each one.

  1. 01Talk
  2. 02Fix
  3. 03Agent
  4. 04Herdr

Hold Talk and speak

Hold Talk on your phone and say it. Your words show up as you speak and are typed at the caret when you let go.

Talkhold to dictate

Coding agents

Manage your agents from the couch.

Vispr speaks Herdr, the terminal workspace for coding agents. It knows which pane you're talking to and who's waiting on you.

  • Pane-targeted dictation. Words go to the Herdr pane that was focused when you started talking. If it changes, Vispr stops rather than guessing.
  • Every tab, one tap away. Your phone shows the workspace's tabs with each agent's state. Tap one to switch, + for a new tab, or hold a tab to rename or close it.
  • Jump to the agent that needs you. Tap the waiting chip to focus the next agent blocked on an approval, then the ones that finished.
  • Steer without typing. Enter sends, Ctrl ⏎ steers an agent mid-task, and Esc interrupts.
  • On a DualSense, the D-pad moves between panes, L3 jumps to whoever's waiting, and Create chords open tabs and workspaces or run /usage and /compact.

Prefer a controller?

A DualSense, if you have one.

Everything on the phone has a button here too, plus sticks under your thumbs and the controller's own microphone. It works on its own or alongside your phone. Pick a mode to see what each control does, or watch the DualSense film.

      Hover a control to find it. Options opens Vispr's settings from anywhere; Create + Options puts it in standby.

      Under the hood

      How it works.

      Everything between your voice and the caret runs locally, in Rust. The input loop never waits for the models.

      1. 01

        Microphone

        Your phone's, the DualSense's or your computer's, resampled to 16 kHz in 20 ms batches.

      2. 02

        Nemotron 3.5

        NVIDIA's streaming speech model, through transcribe.cpp.

      3. 03

        S1-mini

        Superwhisper's dictation-cleanup model, reading ahead as you speak.

      4. 04

        Typing

        A persistent virtual keyboard and pointer on Linux, Accessibility events on macOS. One paste per dictation.

      Focus

      Focus checks

      On Hyprland and macOS, Vispr follows window focus. If you switch away mid-sentence, it cancels instead of typing into the wrong window.

      Room

      Background noise

      Learns your background noise from loudness alone, so the voice meter moves for you and not the fan.

      Mapping

      A tested mapping

      Every control rule lives in one pure, unit-tested mapper, so the same input always produces the same actions.

      Privacy

      Everything stays on your computer.

      No audio uploads, no cloud transcription, no account. While Vispr is awake it reads only the microphone's loudness, to measure background noise. It listens to your voice only while you hold Talk (or R2), and the audio never leaves your computer. This site has no analytics and loads nothing but its own files.

      What do I need?

      A Mac with Apple silicon, or a Linux PC on Wayland: a recent x86-64, or an ARM64 machine such as a Snapdragon laptop. Hyprland (and Omarchy) is the best-tested Linux setup; GNOME and KDE support is new. Then your phone, on the same Wi-Fi. A DualSense over USB or Bluetooth is optional. About 1.2 GB of models download on first run, or none of the speech models if you've used Handy: Vispr reuses its downloads.

      Do I need a controller?

      No. Most people use their phone. Scan the code in Vispr and add the page to your Home Screen: it becomes a trackpad, the keys each mode needs and a button you hold to talk. Voice and controls go straight to your computer over your Wi-Fi. If you have a DualSense, it adds sticks under your thumbs and its own microphone, and the two work together.

      Which languages?

      Nemotron 3.5 transcribes many languages. Smart formatting is English only for now; turn it off, or use Raw mode, for everything else.

      macOS or Windows?

      Vispr runs on Linux and macOS. The macOS version is newer and less tested. Windows isn't planned yet.

      Is it open source?

      It will be, under the MIT licence, when it's ready for a public release. Until then it's in a private alpha.

      Early access

      Try the alpha.

      Vispr is in a private alpha for macOS and Linux. If you have an access code, enter it to download.

      Contribute on GitHub Coming soon