Repository navigation
VocalCode
Free, local-first dictation and meeting notes, open source under
AGPL-3.0-only. There is nothing to buy: no purchase, account or activation
code. Recognition runs on your CPU after the selected model downloads.
New in 1.4.1
- One app, free for good. The early free builds (VocalCode Community 1.3.1
and 1.4.0) and the paid releases are now one app, just VocalCode. A paid
installation updates to it in place and keeps everything (see below). - Setup in three steps: pick your language (its model starts
downloading), check your talk key (add another if your keyboard has no
Right Ctrl), then try a sentence in a box that shows what VocalCode heard. - New installs start with safer defaults: the mouse Back button is no
longer Enter (tick it in setup or Shortcuts if you want it); the speech
gate is on, so a noise-only press types nothing; History is kept on disk,
encrypted, for a week (change it on the History page). Existing installations keep
their current settings. - Notices while you dictate. Pressing the talk key while the model is
still downloading, the microphone is broken or no text field has focus now
shows a short line at the indicator instead of doing nothing. - Model downloads resume after a dropped connection or quitting, instead
of starting over, and a failed download says why and retries. - English defaults to Parakeet on computers fast enough for it (better
with coding words than SenseVoice); smaller computers keep SenseVoice. - Faster in noisy rooms: pauses are found against the room's own noise
floor, so long dictations are prepared while you talk rather than all after
you let go. - Chinese punctuation in mixed Chinese/English sentences stays
full-width; a result that is only punctuation (a stray "。" over noise) is
treated as nothing heard. - Qwen3-ASR: dictations longer than about 38 seconds are no longer cut
off or replaced by the word "language"; long input is decoded in pieces,
and text invented over silence or hum is dropped. - About & help replaces the old licence page, with Copy diagnostics
(nothing is sent). If VocalCode closes unexpectedly, the next start says so
and points to the log. - Tested with speech: the release voice corpus now has 1,199 clips,
including mixed Chinese/English, noise-only presses and 30–90 second
dictations, replayed through the production pipeline on four model routes.
Downloads
- Windows x64:
VocalCodeSetup.exe(publisher: Daming Wu). - Apple-silicon macOS:
VocalCode-<version>.dmg(signed and notarized). - Exact Corresponding Source:
VocalCode-source-<version>.tar.gz. SHA256SUMS: hashes of release assets;latest.json: the app's update feed.
Installing over an earlier VocalCode
VocalCode is now a single free app. It installs where the paid VocalCode
releases (1.2.1 and earlier) were installed and uses the same data folder, so
it replaces such an installation in place: settings, dictionary, snippets,
meetings and downloaded models stay where they are, and nothing needs
importing. A paid release whose licence or trial is still active offers this
version as an in-app update; otherwise run the installer over it (on a Mac,
replace VocalCode.app in Applications). The paid licence and trial files are
left where they are, and VocalCode never reads them. Uninstalling VocalCode
later keeps your data folder.
A paid VocalCode installed with Scoop cannot update itself this way: run
scoop uninstall vocalcode, then install from this page.
The early free builds, VocalCode Community 1.3.1 and 1.4.0, were a separate
app. On Windows, the installer uninstalls VocalCode Community 1.3.1 or 1.4.0
and keeps its data folder; Settings → System → Previous VocalCode copies
what you choose from it. On a Mac, remove VocalCode Community 1.3.1 or 1.4.0
yourself. Those builds cannot update to this version in the app: download it
from this page once. Don't run two copies at once: each would type every
dictation.
Checks
Before publication, every release is tested for installation, upgrade,
signatures, notarization, native linkage and data preservation, and on
Windows for an in-place update of a real paid VocalCode 1.2.1 through its own
updater. Actual microphone, system-audio, meeting and accessibility behavior
still depends on the device and operating-system permissions; these checks do
not promise perfect speech recognition or replace broad native-device testing.
VocalCode updates from this project's GitHub Releases and verifies the exact
size, SHA-256, publisher signature and product identity before installing.
中文
1.4.1 新增：早期免费版（VocalCode Community 1.3.1 和 1.4.0）与付费版合并为同一个应用 VocalCode，付费版可原地更新、数据不动；
首次设置改成三步（选语言、确认说话键、试说一句）；新安装默认：鼠标后退键不再当回车、语音闸门开启
（只有噪声时不输入）、历史记录加密保存在本机一周，已有安装保持原设置；模型没准备好、麦克风异常或没有
输入框时，指示器会给出提示；模型下载可断点续传；性能足够的电脑上英语默认用 Parakeet；嘈杂环境里更早
开始识别；中英混说时中文标点保持全角，噪声里只识别出一个标点时不再输入；Qwen3-ASR 超过约 38 秒的
长听写不再被截断或变成 "language"；新增「关于与帮助」和意外退出提示。发布前用 1199 段合成语音
（含中英混说、纯噪声、30–90 秒长听写）在四条识别路线上做了完整回放测试。
VocalCode 现在是一个免费应用，按 AGPL-3.0 开源，无需购买、账户或激活。下载：Windows 用
VocalCodeSetup.exe，Apple 芯片 Mac 用 VocalCode-<版本>.dmg，源码为
VocalCode-source-<版本>.tar.gz。两个平台的安装包均经过签名，Mac 包完成 Apple 公证。
VocalCode 安装在付费版（1.2.1 及更早）原来的位置，使用同一个数据文件夹，所以会原地替换付费版：
设置、词典、片段、会议记录和已下载的模型原样保留，无需导入；付费版的授权与试用文件保持原样，
VocalCode 从不读取。授权或试用仍有效的付费版会在应用内提示这次更新，其他情况请直接覆盖安装
（Mac 上替换“应用程序”里的 VocalCode.app）。之后卸载 VocalCode 也会保留数据文件夹。
通过 Scoop 安装的付费版无法这样更新：请先运行 scoop uninstall vocalcode，再从本页安装。
早期免费版 VocalCode Community 1.3.1 和 1.4.0 是单独的应用。在 Windows 上，安装程序会卸载
VocalCode Community 1.3.1 或 1.4.0 并保留它的数据文件夹，可在 设置 → 系统 → 旧版 VocalCode
里选择要复制的内容；在 Mac 上请自行删除 VocalCode Community 1.3.1 或 1.4.0。这两个早期版本无法在
应用内更新到本版本，请从本页下载一次。不要同时运行两个副本，否则每次听写都会被输入两遍。
首次下载模型和检查更新需要联网；录音与识别在本地处理。
Maintainer: Daming Wu.
Issues: https://github.com/wudaming00/vocalcode-community/issues