Qugora is in pre-release: downloads, sign-in and Pro subscriptions aren’t open yet.

System-wide voice input for macOS and Windows

Voice input where fixingbeats the keyboard

Speech-to-text is a solved half. Qugora finishes the other half: when a number, an English word or a name comes out wrong, swipe across it on the trackpad, pick a candidate, or just re-say those few words. Your hands never go back to the keyboard.

macOS · coming soon Windows · coming soon

Pre-release: installers aren’t available yet. Downloads will appear here at launch.

Free to use. New accounts get a 14-day Pro trial. Audio stays on your device by default.

Let’s sync with the chat gbt team next Wednesday about the 3 hundred k budget.

Drag across (or tap) a dotted word to fix it yourself.

01Speak naturally

Talk the way you think. Get finished text.

Pause, say “um”, change your mind halfway through. When you stop, a model drops the filler, adds punctuation, keeps what you meant, and types it at your cursor.

  1. What you said

    Um, so, let’s, uh, sync with the chat gbt team next Wednesday about the, like, 3 hundred k budget.

    fillers, pauses, restarts

  2. After AI cleanup

    Let’s sync with the chat gbt team next Wednesday about the 3 hundred k budget.

    fillers gone, punctuation in

  3. Still wrong

    Let’s sync with the chat gbt team next Wednesday about the 3 hundred k budget.

    English, numbers, names: cleanup can’t guess them

But English terms, numbers and names are guesses cleanup can’t make. That last step is what swipe, pick and re-say are for.

02Swipe to fix

A few words wrong? Swipe, pick, or re-say.

This is where Qugora differs from other voice input: fixing doesn’t send you back to the keyboard either. Find the cursor, delete, retype becomes one finger on the trackpad.

  1. 1 / 3

    Swipe the wrong words

    Drag left-to-right across them on the trackpad, like circling them with a finger. Close enough is fine.

  2. 2 / 3

    Pick a candidate

    A candidate bar pops up. The model uses context and your personal lexicon, so the first one is usually right.

  3. 3 / 3

    Or just re-say them

    None right? Say those few words again, or describe them (“Lin, as in two trees”), and only that span is replaced.

03Every app

In the text fields you already use

Qugora is a system-level input layer: speak, and the text lands at your cursor. To those apps it’s just text appearing in the box. They need no integration and never know Qugora is there.

  • WeChat
  • Feishu / Lark
  • DingTalk
  • ChatGPT
  • Grok
  • Email
  • Docs
  • Notes
  • Browser forms
  • Anywhere you can type

04Personal lexicon

What you fix, it remembers

Every swipe-fix saves the corrected spelling to an on-device lexicon: colleagues’ names, project codenames, English jargon, number formats. Recognition and candidates prefer them next time, so the first candidate is right more and more often.

Fix onceSaved to lexiconFirst choice next time

Personal lexicon

  • ChatGPT chat gbt term
  • Lin Wei Lynn Way name
  • $300k 3 hundred k number
  • Q1 OKR Q one OK project

Stored only on your computer. Edit or delete any time.

05Confirmed memory

It only remembers what you approve

Confirmed text can become local entries: meetings, to-dos, memos, ideas. Each one stays only if you tap “Keep”; “Discard” throws it away. Only approved items sync, and only if you turn sync on (Pro). It’s how a future personal agent remembers you, and it stays yours.

Awaiting your approval

  • Meeting Budget sync with the ChatGPT team, Wed
  • To-do Send the weekly report to Lin Wei
  • Idea Remember candidate preferences per app

Try Keep or Discard.

06Privacy

It stays on your computer.

Audio and text stay on your computer by default. Only what you approve ever leaves it.

  • Audio and text stay on device

    Speech recognition runs locally. What you say and type stays on your computer by default.

  • Memory needs your OK

    Memory items are created locally and need your approval one by one. Only approved items sync — and only if you turn on sync (Pro).

  • Telemetry is counts and timings only

    To improve correction we collect counts and timings (e.g. “was the first candidate right?”) — never text. You can turn it off any time.

  • Cloud LLM requests aren’t stored

    With Qugora Cloud we record token counts for billing and nothing else. Or bring your own model key.

Read the privacy policy (draft)

How it compares

Everyone turns speech into text. Qugora adds fixing and memory.

Everyone does the first two rows well. Qugora mostly adds the rest.

QugoraWispr FlowTypelessBuilt-in OS dictationDoubao IME / Shandianshuo
Dictate into any app’s text field Yes Yes Yes Yes Yes
AI cleanup (filler words, punctuation) Yes Yes Yes Basic punctuation Yes
Swipe wrong words on trackpad → candidate bar Yes Not publicly documented Not publicly documented Not publicly documented Not publicly documented
Re-say only the selected words Yes Not publicly documented ¹ Not publicly documented Not publicly documented Not publicly documented
Personal lexicon / dictionary Learns from every swipe-fix Personal dictionary (can add words automatically) Personal dictionary Not publicly documented IME user dictionary
Where audio is processed by default On device (local ASR); cloud LLM optional Cloud (see their privacy policy) See their privacy policy On-device or cloud, depending on OS and language See their privacy policy
Local memory library (you approve each item) Yes Not publicly documented Not publicly documented Not publicly documented Not publicly documented

Based on each product’s public information as of November 2026; things may have changed. “Not publicly documented” means we couldn’t find it, not that it doesn’t exist. Spot an error? Email hello@qugora.com and we’ll fix it.

¹ Some products let you transform selected text with a voice command (e.g. “make this more formal”). That’s different from re-saying a few words to correct a recognition error.

FAQ

Which apps does it work in?
Pretty much any text field: WeChat, Feishu/Lark, DingTalk, ChatGPT (web and desktop), email, docs, browser forms. Qugora works as a system-level input layer and types at your cursor, so apps need no integration.
No trackpad — can I use a mouse?
Swiping is just selecting text, so a mouse drag also brings up the candidate bar. We design for the trackpad because keeping your hand on it is smoothest. The keyboard is always there as a fallback.
Does it need internet?
Speech recognition runs on device, so basic input works offline. AI cleanup and candidates need a model: use Qugora Cloud, bring your own API key, or connect a local model.
How well does it handle mixed Chinese/English, numbers and names?
That’s exactly where Qugora focuses. First-pass recognition will make mistakes, so we prioritize making fixes fast — and every fix teaches your lexicon, so the same word is more likely right next time.
Free vs Pro?
Free includes local recognition, local correction and the local memory library, works with your own model key, and comes with a small monthly Qugora Cloud allowance. Pro adds a much larger cloud LLM quota and memory sync. New accounts get a 14-day Pro trial — see Pricing.
Is my data used to train models?
No. Audio and text stay on your computer by default; telemetry is counts and timings only, never text; for cloud LLM requests we only record token counts.
How is this related to Voconly?
Qugora is built on Voconly, an MIT-licensed open-source project, adding swipe-to-correct, the personal lexicon and the memory library on top of its system-wide voice input. Thanks to Voconly’s author and contributors.

Next time a number comes out wrong, just swipe it.

Free download. New accounts get a 14-day Pro trial.

macOS · coming soon Windows · coming soon

Pre-release: installers aren’t available yet. Downloads will appear here at launch.