Open-source dictation that types at your cursor in any app. Recognition runs offline with a local model, or on the built-in service or your own provider; AI polish tidies the wording with the preset you choose.
Hold Ctrl+Alt+Space and speak; the overlay shows what you say
Release the keys; the text is proofread and inserted at your cursor
The review is on Monday at two. Please bring the latest numbers.
Each step has its own settings, and AI polish can be turned off. If the dictionary, AI polish or the rules fail, the dictation completes with the text from the step before.
Hold to talk, or press once to start and again to stop; record the microphone, the sound the computer plays, or both.
A local model, the built-in service or your provider, on the whole recording or as you speak. Your dictionary's right spellings and the scene's terms go with the recording as hotwords, so cloud recognition keeps them; local models take none.
After recognition, what a word is heard as is replaced with its right spelling, for local models and cloud services alike.
Tidies the text with the current preset or scene: filler words go, punctuation comes in, corrections are kept.
Rewrite the final text in list order, by a literal match or a regular expression.
Pastes at your cursor and puts the clipboard back, keeping a copy in the history on this computer.
The whole process is described on What is Voltip.
AI polish removes filler words, fixes the punctuation and keeps the correction you made, without rewording what you said. You can turn it off.
Um, the review is on Monday at ten, no, sorry, at two. Uh, please bring the latest numbers.
The review is on Monday at two. Please bring the latest numbers.
um so I think we should uh ship the fix on friday and then update the docs next week
I think we should ship the fix on Friday and then update the docs next week.
Turn it into a bulleted list.
Selected textLet's meet on Thursday at three in the main conference room to go over the release plan.
The review is on Monday at two. Please bring the latest numbers.
- Let's meet on Thursday at three in the main conference room to go over the release plan.- The review is on Monday at two. Please bring the latest numbers.
“Actual output” comes from a screen recording of Voltip. “Example” is illustrative; real results depend on the AI model you choose.
Eight are built in, with their prompts open to read, and any of them can be copied into a preset of your own and changed.
um so the thursday release let's push it to next monday and uh testing still has two open issues I'll let everyone know once they're fixed
ProofreadFixes typos and punctuation, removes filler words
So the Thursday release, let's push it to next Monday, and testing still has two open issues. I'll let everyone know once they're fixed.
Clarify intentKeeps the corrections, lists several points
- Push the Thursday release to next Monday.- Testing still has two open issues; let everyone know once they're fixed.
FormalComplete, formal written language
The Thursday release will be postponed to next Monday. Testing still has two open issues, and I will inform everyone once they have been resolved.
Chinese ⇄ EnglishFixes recognition errors, then translates
周四的发布推迟到下周一。测试还有两个问题待解决，修复后我会通知大家。
Example outputs; real results depend on the AI model you choose. The others are Prompt optimizer, Casual chat, Key points and Punctuation only.
Built-in scenes apply once you turn them on; you can add a scene for any app.
VS CodeAnd Windows Terminal, PowerShell and each system's own terminal.
Coding agents in the terminal
And Claude Code and other coding agents that run in a terminal.
Keeps code, commands, paths, identifiers and Markdown as they are.
%20--%3e%3csvg%20id='SVGRoot'%20width='24px'%20height='24px'%20version='1.1'%20viewBox='0%200%2024%2024'%20xml:space='preserve'%20xmlns='http://www.w3.org/2000/svg'%20xmlns:xlink='http://www.w3.org/1999/xlink'%3e%3cdefs%3e%3clinearGradient%20id='linearGradient4349'%20x1='3'%20x2='21'%20y1='23'%20y2='1'%20gradientUnits='userSpaceOnUse'%3e%3cstop%20stop-color='%23083fa6'%20offset='0'/%3e%3cstop%20stop-color='%232dcbfd'%20offset='1'/%3e%3c/linearGradient%3e%3c/defs%3e%3cg%20fill-rule='evenodd'%20stroke-linecap='square'%3e%3cpath%20d='m5%200c-1.662%200-3%201.338-3%203v18c0%201.662%201.338%203%203%203h14c1.662%200%203-1.338%203-3v-12l-9-9h-8zm10%200%207%207v-7h-7z'%20fill='url(%23linearGradient4349)'/%3e%3cg%3e%3cpath%20d='m2%2020v1c0%201.662%201.338%203%203%203h14c1.662%200%203-1.338%203-3v-1c0%201.662-1.338%203-3%203h-14c-1.662%200-3-1.338-3-3z'%20fill-opacity='.19965'/%3e%3cpath%20d='m15%200h1.4l5.6%205.6v1.4z'%20fill-opacity='.19965'/%3e%3c/g%3e%3cg%3e%3cg%20fill='%23fafafa'%3e%3crect%20x='6'%20y='9'%20width='5'%20height='1'/%3e%3crect%20x='6'%20y='11'%20width='5'%20height='1'/%3e%3crect%20x='6'%20y='13'%20width='5'%20height='1'/%3e%3crect%20x='6'%20y='15'%20width='12'%20height='1'/%3e%3crect%20x='6'%20y='17'%20width='5'%20height='1'/%3e%3c/g%3e%3cpath%20d='m12%209v5h6v-5h-6zm4.5%201a0.5%200.5%200%200%201%200.5%200.5%200.5%200.5%200%200%201-0.5%200.5%200.5%200.5%200%200%201-0.5-0.5%200.5%200.5%200%200%201%200.5-0.5zm-2.5%201%202%202h-3l1-2z'%20fill='%239ae9f9'/%3e%3c/g%3e%3c/g%3e%3c/svg%3e){kind=small}
LibreOffice WriterAnd Outlook, Word, WPS, OneNote, Typora, and Mail and Pages on a Mac.
Complete sentences and natural paragraphs for mail and documents, with no invented greetings or facts.
And WeCom, Slack, Teams, DingTalk, Feishu, Lark, and Messages on a Mac.
No extra instructions: the Casual chat preset writes short, conversational sentences without a closing full stop.
Each field is a scene of its own and applies once you add the apps you use.
Strict proofreading that keeps specialist terms as they are and writes numbers, units and dates correctly.
On this computer, with the built-in service or with your own provider; switch at any time on the Speech models page.
Uses a model downloaded to this computer
The audio stays on this computer
Models
Run by the Voltip project
The server keeps no recordings or text
AI polish models
Uses your own API key
The recording goes straight to the provider you chose
Recognition providers
After recognition, a language model tidies the text. It can run on the built-in service, your provider, or Ollama on your own computer.
On an NVIDIA L40S, Qwen3-ASR 0.6B recognised 16 seconds of Chinese speech in 0.16 s on the GPU and in 2.87 s on 8 processor threads. See Local recognition and Cloud services.
These are in the desktop app; the ones that start off can be turned on in Settings.
Add a word the recogniser gets wrong once, and it is spelled right from then on; cloud recognition and AI polish get the right spelling as a hint too. Rules rewrite the text last, with regular expressions if needed. In the history, select a misheard word to add it.
cube controlkubectl
Right Ctrl, Right Alt, Fn on a Mac, or the middle or side mouse buttons can start dictation on their own. Pressing another key while you hold one cancels the recording, so Right Ctrl + C keeps working.
Right CtrlRight AltFn
Record the microphone, the sound the computer plays, or both, with the speakers' echo removed when mixing. A recording can run up to 2 hours and be exported as SRT subtitles or text.
Recognition and AI polish each keep a list of up to 8 fallback models. When the selected model's quota is used up, Voltip moves on to the next and the dictation still goes through; other errors, such as a lost connection or a wrong key, are reported and never switch models.
Claude Code, Codex and other AI tools can read the history, dictionary and rules over MCP, collect the misrecognitions that keep coming back, and add entries and rules with your permission. Every change can be undone.
Other programs send a recording to an OpenAI-compatible API and get back text processed by the same recognition, dictionary, presets and rules. Turn it on in the app, or run voltip-server on a Linux server without a desktop; by default it accepts connections from this computer only, and every request needs an access token.
voltip-server --scene coding --preset promptPair an Android phone with the computer and hold to talk on the phone. The computer recognises the speech with its own settings and inserts the text at its cursor. With no computer online, the phone transcribes on its own.
Scan the QR code on the computer and confirm on the phone; with the 6-digit code, also check that both screens show the same safety code.
The audio streams to the computer, compressed with Opus and encrypted end to end. The computer recognises it and inserts the text at its cursor.
Type on the phone or send its clipboard, and the computer inserts it like a dictation result.
The computer's history and settings are synced to the phone, to read there.
The devices connect through a relay that forwards encrypted data only, and reconnect on their own after a change of network.
The desktop app is the same on all three systems; the Android app makes a phone the computer's microphone and keyboard.
irm https://raw.githubusercontent.com/sunerpy/voltip/main/scripts/install.ps1 | iexcurl -fsSL https://raw.githubusercontent.com/sunerpy/voltip/main/scripts/install.sh | shOne command picks the right package for the computer, checks it against the release's SHA256SUMS and installs it. Packages for every platform and their checksums are on the releases page; the install guide covers each platform.
An iOS app is planned for a later release. Packages for Windows on ARM and Linux on ARM are not planned for now.
More questions and troubleshooting are on the Questions page.
Yes, with a local model. Download it once; recognition then works offline. AI polish needs a service, unless you run Ollama on your computer.
No. The microphone opens when you press the shortcut and closes when the recording ends.
It depends on the services you choose. Local recognition sends nothing; cloud recognition sends the recording to the built-in service or the provider you set up; AI polish sends text only, never audio. See Privacy.
No. The Voltip project runs it, all users share it, and requests are limited per address. For AI polish you can rely on, use a provider of your own.
Choose how Voltip connects under Settings → Network. It follows the system's proxy settings by default, or uses a proxy server you enter. Recognition, polish, model downloads, updates and the connection to your phone all follow this setting.
No. Choose Press to start, again to stop, or Hold or press, where a short press locks the recording. A recording stops on its own at the length you set; Voltip has no wake word and does not stop when it hears silence. See Shortcuts and recording modes.
The Windows installer is not code-signed yet, so SmartScreen warns on the first start. The macOS app is signed with the project's own certificate but not notarized by Apple, so you confirm it once on first opening. Every release comes with SHA256SUMS and build attestations, and the one-line install checks the download against SHA256SUMS. See Install.
An iOS app is planned for a later release. An Android app is available now.