Accessible Live Captions v0.5.0

Accessible Live Captions 0.5.0

Live captions for anything your PC can hear — built so that a screen-reader user can
actually read back what was said, not just glimpse it.

Windows has Live Captions, but it shows a sliver of text that vanishes, keeps no
history you can review, and is effectively invisible to Narrator, NVDA, and JAWS. This
app captions the same audio — a meeting, a video, one specific application, or your
microphone — entirely on your own machine, and puts every finished line into a full,
scrollable transcript where each line is its own focusable item you can arrow through
at your own pace.

Everything runs on-device. No cloud, no account, no API key, and no audio ever
leaves your PC.

New in 0.5.0: an experimental x64 build

Every previous release was ARM64-only, which left everyone on an Intel or AMD PC at a
dead end. This release adds a portable x64 build so the app can at least be tried
on those machines.

Please read this before downloading it. The x64 build is genuinely experimental:

• It is built and architecture-verified by CI, but it has never been run on a real
  x64 machine. I do not own one, so nothing about it has been smoke-tested — not
  audio capture, not recognition, not the NPU. It may not work at all.
• Do not run the x64 build on an ARM64 PC. Windows will happily run it under x64
  emulation, and in the one instance it was tried that way the machine hung hard enough
  to require a hardware-watchdog reset. That is an emulation-layer problem and should
  not affect a real x64 machine, but there is no reason to take the risk: on ARM64,
  use the ARM64 download.
• There is no x64 .msix, so the Windows on-device (NPU) engine is not available
  on x64. The other three engines are all present. An installer that cannot be tested
  is worse than no installer.
• If it works — or if it doesn't — please
  open an issue.
  A report from an actual x64 machine is the only thing that can move this from
  experimental to supported.

ARM64 users are unaffected: the ARM64 MSIX and portable zip are built exactly as
before, from the same code, and remain the verified builds.

Which download?

File	What you get
.msix (ARM64)	The full app, including the Windows on-device (NPU) engine. Double-click to install — it is code-signed, so no Developer Mode and no certificate steps. This is the recommended download on a Snapdragon PC.
-win-arm64-portable.zip	Unzip and run; nothing is installed. Handy from a USB stick or where you can't install software. No NPU engine, which Windows grants only to installed apps.
-win-x64-portable-EXPERIMENTAL.zip	Intel/AMD. Unzip and run. Untested — see the warnings above.

Two ways to see captions

Press F7, or use View ▸ Presentation, to switch at any time.

• Transcript — the full scrollable history. Arrow keys, Home and End move
  through it line by line; new captions never steal focus while you are reading.
• Panel — one caption at a time in large text, the way the built-in Windows Live
  Captions presents itself.

Choosing the compact panel costs you nothing. Every caption is still kept, so switching
back to Transcript shows the whole session from the first line, and captions are
announced to your screen reader identically in both. In panel mode, ↑ and ↓ step
back through earlier captions and pause following so an arriving caption can't pull you
off the line you're reading; End or Follow live resumes. The readout tells you
which state you're in — "Live — 24 captions" or "Caption 6 of 24 — paused".

Four speech engines

Choose under Audio ▸ Engine. All of them run on your machine.

Engine	Words appear	Text style	Notes
Whisper (default)	After each phrase	Punctuated	Model picker from Tiny to Large v3; downloads on first use
Streaming	Word by word, ~½ s behind	Lowercase, no punctuation	Lowest delay and very light on the CPU; 456 MB model on first use
Windows on-device (NPU)	After each phrase, fast	Punctuated	The recognizer Windows itself uses. Runs on the neural processor, so it costs almost no CPU or battery. English only; ARM64 .msix only
Windows speech	Instant	Plain	No download, but microphone only

What you can caption

Your microphone, everything the PC is playing, or one specific application
(useful for captioning a meeting without music or notifications bleeding in).

If you use a screen reader, Audio ▸ Exclude screen reader speech — on by default —
keeps NVDA, JAWS, or Narrator out of the captions, so a transcript of a meeting isn't
interleaved with your own screen reader's narration.

Reading and keeping captions

• Announcements of new captions can be toggled with F8.
• Caption text scales from 14 to 48 point with Ctrl + and Ctrl −, and the app uses
  system colors so High Contrast themes apply automatically.
• Optional timestamps, Copy all (Ctrl+Shift+C), and Save transcript (Ctrl+S).
• Tab deliberately does nothing anywhere in the app — the transcript is the content,
  and every command lives on the menus, reached with Alt or F10.

Installing without the mouse

Add-AppxPackage -Path .\AccessibleLiveCaptions-v0.5.0-win-arm64.msix

If the App Installer dialog sits at 0% with no error, the problem is Windows'
package-deployment queue rather than this download — a reboot clears it. A healthy
install of this package takes under a second, since it has no downloadable
dependencies, so any wait beyond a few seconds means something else is holding the
queue. The
project status document
explains how to confirm that from the event log.

Known limitations

• The x64 build is unverified. See the section above. ARM64 remains the only
  tested architecture.
• The NPU engine uses a preview Windows API: English only, ARM64 MSIX only, and a
  future Windows update could change its behavior. The other three engines are
  unaffected.
• Per-application capture and screen-reader exclusion don't work on audio devices that
  don't support per-process capture. The app detects this within a few seconds, tells
  you, and keeps captioning everything rather than silently going quiet.
• Windows SmartScreen may warn on first download until the release builds up
  reputation. Choose More info, then Run anyway.
• The first start of the Whisper or Streaming engine downloads its model; the status
  line reports progress.

Full documentation is in the
user guide.