Speak naturally. Keep your work local.
CPU-first dictation and meeting notes for Windows and Apple-silicon Mac.

English · 简体中文

Features · Getting started · Build from source · Privacy · Maintainer & contact · Contribute

Maintained by Daming Wu · AGPL-3.0-only

VocalCode is free and open source (AGPL-3.0): every feature, no account, no activation. Downloads and release status include signed Windows installers and notarized Apple-silicon DMGs only after the release checks pass. Tags named source-preview-* contain source only. See validation scope and limitations.

Why I'm building this

I'm Daming Wu, the maintainer of VocalCode. I started with a simple need: speaking longer prompts into coding tools instead of typing every word. The project now also explores local meeting notes, reusable vocabulary, and recoverable dictation history.

I want recognition to run on my own machine, with CPU-friendly options and clear controls over what is recorded and retained. Opening the source is a way to develop that in public: reproducible bug reports, real language feedback, and improvements that others can inspect and share.

Help shape the original project. Multilingual test cases, microphone and echo regressions, keyboard accessibility, and small well-tested fixes are especially useful. You do not have to be a Rust developer to contribute. See the near-term priorities, how to contribute, and who maintains it / how to get in touch.

What you can do

Workflow	What VocalCode does
Dictate into your apps	Hold a shortcut, speak, and insert text into a supported focused field. Focus checks protect against accidental delivery to the wrong target.
Take local meeting notes	Record microphone and system audio, import audio, search transcripts, add bookmarks and speaker labels, and export notes. Recording always needs your confirmation and participants' permission.
Teach your vocabulary	Add dictionary entries or review a learned correction, edit it, or undo it.
Recover your words	History keeps recent dictations encrypted on this device (Off / 24 hours / 7 days); text that could not be typed stays there until you quit. Optional local diagnostics have their own storage controls.
Choose your trade-offs	Select local language/model routes, the non-speech noise filter, and experimental pause-delimited progressive typing. No GPU is required.

This is a desktop tool, not a meeting bot or a cloud transcription subscription.

Defaults for a new install

Setting	New install	Where to change it
Mouse Back button (X1) sends Enter	Off. First run offers it as an unticked box, Use the mouse Back button as Enter.	Shortcuts → Tap to send
Filter non-speech noise	On. Without it, fan, keyboard or pink noise can be typed as "I.", "그." or "我。".	Settings → Dictation
Keep history	7 days: up to 50 recent dictations, encrypted with your Windows account or a Keychain key, then deleted. Off deletes what was kept. Uninstalling does not; Remove… under Settings → System does.	History

Updating does not change these for an existing install. Settings written by 1.4.0 or earlier keep the Back button as Enter if they had it, keep the noise filter off and keep History to the current session, until you change them. Settings saved by this version use a newer settings format that 1.4.0 and earlier will not start with, so going back to an older release is not supported.

The current development tree also includes an opt-in Windows desktop control bar and a manual rewrite scratchpad. The bar is off by default and uses the existing dictation engine; rewriting defaults to an already-installed local Ollama model, with a separately consented Claude CLI option. These changes have not been published as a new release by this work. See the development evaluation and rewrite boundaries.

Install or build

Open Releases and choose a stable v* release: VocalCodeSetup.exe for Windows x64 or VocalCode-<version>.dmg for Apple-silicon macOS. Each includes matching source and SHA-256 checksums. On macOS, drag VocalCode.app to Applications.

Install the native prerequisites in BUILDING.md, then:

git clone https://github.com/wudaming00/vocalcode-community.git
cd vocalcode-community
cargo build -p vocalcode-app --release --locked --features community

No payment account, activation code, or operator credential is required. The community build feature (on by default) is the free build; the explicit flag only makes that clear. A build from source is a development build: it keeps its own data folder (VocalCode Dev) and login item, so it never reads or changes an installed VocalCode's data. See BUILDING.md.

Coming from an earlier VocalCode. VocalCode installs where the paid VocalCode releases (1.2.1 and earlier) did and uses the same data folder, so installing it over one of them replaces that app in place and keeps its settings, dictionary, meetings and downloaded models; nothing needs importing. If you installed the paid VocalCode with Scoop, run scoop uninstall vocalcode first (your data folder stays), then install VocalCode: a copy Scoop unpacked cannot update itself in place, and its update attempt fails without changing anything. Turn on launch at login again in VocalCode if you used it. The early free builds, VocalCode Community 1.3.1 and 1.4.0, had their own installation: the Windows installer uninstalls that app and keeps its data folder, and Settings → System → Previous VocalCode copies what you choose from it. Don't run two copies at once: each would type every dictation.

Getting started

First run takes three short steps:

1. Pick the language you speak most. Only that language's on-device model is downloaded, and the download starts right away.

2. Check your talk key. VocalCode records only while a talk key is held. These are bound out of the box:

   Platform	Default talk keys
   Windows	Mouse forward button (X2), Right Ctrl
   macOS	Mouse forward button (X2), F13, Right Option

   Many laptop keyboards have no Right Ctrl. If yours doesn't, add another key in this step or later under Shortcuts.

   The same step offers Use the mouse Back button as Enter, unticked. Ticking it makes Back send what you dictated, and other apps stop receiving it as Back.

3. Try it. Once the model is ready, hold the key, say a sentence, and let go: the text appears in the box. Skip the step if the download is still running.

After that, the text goes into whichever text field has focus.

Privacy and network access

Speech recognition and meeting audio processing run on the device after the selected model is downloaded. There is no account, telemetry, analytics, licence check or checkout. Microphone audio, transcripts, history, your dictionary and meeting notes are not uploaded. These are the only network connections the app makes:

When	Connects to	What is sent
A speech model (or its punctuation model) is first needed	models.vocalcode.app	Requests for pinned model files. Each file's size and SHA-256 are checked before use.
At startup, every 6 hours, and when you press Check now	GitHub Releases of this repository	A request for latest.json. An installer is downloaded from GitHub only after you choose Update, and its size, SHA-256, publisher and product name are verified before it runs.
Only if you connect Upcoming calendar meetings (Beta)	Google (accounts.google.com, oauth2.googleapis.com, www.googleapis.com)	Your Google sign-in and read-only requests for upcoming event metadata, directly between this device and Google. Audio and notes are never sent.
Only if you use the rewrite scratchpad with Ollama	127.0.0.1:11434 on this computer	The text you put in the scratchpad, to your own local Ollama. It does not leave the device.
Only if you choose Claude Code in the scratchpad and consent for that request	Whatever service your Claude Code CLI is configured for	The source text of that one request, through the separately installed CLI. Its service, account policies and logging are a separate trust boundary; a locally installed CLI is not local inference. Codex is detected only, not used.
Only when you click a link in About & help	github.com, in your browser	Nothing from VocalCode; your browser opens the page.

• Copy diagnostics (About & help) puts the version, OS, hardware summary, model, microphone name and the last 200 lines of vocalcode.log on your clipboard. Nothing is sent; you decide where to paste it. The log never contains transcripts, and your home-folder path is shortened to ~.
• Recording, clipboard history, exports, kept History and optional diagnostic persistence can contain sensitive information. Kept History is encrypted and expires, and Remove… under Settings → System deletes it with the rest of the app data; uninstalling does not. Text that could not be typed, including text refused by a password field and copied to the clipboard instead, is kept like any other dictation. Review your OS sync and backup settings. Local speech recognition is not the same as an air-gapped app.

Languages and performance

Different languages use different local model routes, including Parakeet, SenseVoice, and Paraformer; experimental routes are not a blanket quality promise. Model capability is not the same as native-speaker product validation. Latency depends on CPU, model size, utterance length, and capture setup.

See third-party notices and the model integrity manifest. Model weights have their own licences, independent of the client source licence.

Current limitations

• Windows x64 is the local validation environment. GitHub Actions also builds on Apple silicon; a successful hosted build does not replace native-device audio, permissions, and installation tests before a binary release.
• Overlapping speakers and loudspeaker echo can still damage transcription. The meeting echo guard is not a perfect speech-separation system.
• Progressive typing is experimental: it inserts stable segments, not a continuously rewritten live hypothesis.
• Meeting detection is advisory, not proof that a call has started or ended.
• Linux packaging is not provided.
• A paid VocalCode release offers updates in the app only while its licence or trial is active; otherwise install this release over it from Releases. The early free builds (VocalCode Community 1.3.1 and 1.4.0) cannot update to this release in the app: download it from Releases once.

Project layout

vocalcode-app/       Desktop UI, workflows, model management, free-build policy
vocalcode-core/      Portable engine, configuration, and text processing
vocalcode-platform/ Native audio, focus checks, and input delivery
vocalcode-meeting/  Meeting storage, segmentation, echo handling, and export
vendor/             Required patched native bindings with upstream licences

Contributing and licensing

Start with the contribution guide and security policy. The remaining release decisions are explicit in PUBLICATION_BLOCKERS.md.

VocalCode's first-party desktop source is AGPL-3.0-only. Commercial use is allowed under its terms; this is not a noncommercial licence. Redistribution and modified network services carry corresponding-source obligations. Your ordinary recordings and transcripts do not become public by using the app. See LICENSE, licensing scope, and brand guidance. Models and other third-party components retain their own licences.

Support is best-effort, with no promised response time, fix date, or service-level agreement. Existing purchases are a separate matter; this source preparation does not announce a change to their terms. Never put private recordings, credentials, or security exploits in a public issue.