VoxCtrl v0.8.3

What's new in 0.8.3

New: Phonon-2, a more accurate English speech engine for the Parakeet
backend.

In Settings → Engine, with Parakeet selected, you can now pick Phonon-2
(English only, ~735 MB). It is Fermion Research's retrain of Parakeet TDT
0.6B v3 with roughly 2-bit encoder weights, and it averages 5.2% word error
on the Open ASR Leaderboard's English sets — ahead of the Redux variant
(5.7%) and close to the full-size Parakeet v3 (5.0%). Because it only
transcribes English, the Language field is hidden for it. Fermion ships the
model in a format only its own runtime reads, so VoxCtrl downloads the
community ONNX conversion
tiyuvta/Phonon-2-ONNX. It
uses about 1.1 GB of RAM once loaded. Phonon-2 is credited in About
(CC-BY-4.0).

Improved: long recordings with any Parakeet model. Audio longer than 35
seconds is now split into 25–35 second pieces, each cut at the quietest
moment in that range so the cut lands in a pause between words, and the
transcripts are joined. This keeps memory use flat on long dictations; the
speech engine's memory cost otherwise grows with the square of the recording
length.

Improved: model downloads. Model files are now written to disk as they
arrive instead of being held in memory first, and a download that ends short
is discarded instead of being kept as if it had finished.

Notes

Phonon-2 is a third-party conversion, not Fermion's own release. Switching
models downloads the new files; your existing Parakeet models are untouched.

Downloads

Pick the build for your platform and GPU:

Platform	File suffix	GPU required
Linux	-linux-x86_64-vulkan.AppImage	None — GPU used if present
Linux (newer distros)	-linux-x86_64-vulkan-webgpu.AppImage	None — GPU used if present; also accelerates Moonshine and Parakeet. Needs glibc 2.38+ (Ubuntu 24.04, Fedora 39, Debian 13, Arch and similar)
Windows	-windows-x86_64-webgpu.exe	None — GPU used if present

There is one Windows build. It accelerates the Moonshine and
Parakeet speech engines on any Direct3D 12 GPU — NVIDIA, AMD or
Intel — through ONNX Runtime's WebGPU provider, and runs on the CPU
when no usable GPU is found, so it replaces the separate CPU
installer rather than sitting beside it. Each engine has its own
GPU acceleration checkbox in Settings → Engine, and the chip
beside the engine's name says which device it is using. Whisper.cpp still runs on the CPU on Windows either
way, because its only GPU path here is ggml's Vulkan backend, which
does not register on Windows MSVC static builds. If your GPU is not
usable the app says so in the log and falls back to the CPU rather
than pretending. If you were running the separate CPU installer,
VoxCtrl's own updater will move you onto this one; nothing to do.

Linux has two builds. The Vulkan AppImage accelerates
any NVIDIA/AMD/Intel GPU through the host Vulkan driver and runs on
the CPU when there is no Vulkan device, so it replaces the separate
CPU AppImage rather than sitting beside it — the same file is right
whether or not you have a GPU. The Vulkan + WebGPU AppImage is the same
app plus GPU acceleration for Moonshine and Parakeet; it bundles Dawn, which
needs glibc 2.38 or newer, so it will not start on Ubuntu 22.04 or Debian 12 —
use the plain Vulkan AppImage there. VoxCtrl's updater keeps each install on
the variant it is running. If you were running the old CPU AppImage, the
updater moves you onto the Vulkan one; nothing to do.

To keep an engine on the CPU regardless, untick GPU acceleration for it
in Settings → Engine. Worth doing if your only Vulkan device is a software
rasterizer such as Mesa's lavapipe, which is slower than the CPU
backend.

The Linux .deb now requires libvulkan1, which the package
manager pulls in for you.

The Windows build is an NSIS installer, not code-signed yet, so
SmartScreen will warn that the publisher is unknown: choose
More info → Run anyway.

Windows — first run

Windows 10 21H2 or newer, or Windows 11. Everything VoxCtrl needs is
in the installer; WebView2 ships with Windows 11 and the installer
fetches it on older builds.

Two things are worth knowing before the first dictation:

• Microphone access. Windows denies it silently. If dictation
  produces nothing, check Settings → Privacy & security →
  Microphone and allow desktop apps.
• Global shortcuts arrive through a Windows keyboard hook, so
  every gesture works — but Windows does not deliver keys to it
  while an elevated application has focus, or on the UAC prompt and
  lock screen. Shortcuts do not fire there, and dictated text cannot
  be typed into an elevated window unless VoxCtrl is elevated too.

Linux — first run

The Vulkan AppImage needs glibc 2.35 or newer (Ubuntu 22.04+, Linux Mint 21+,
Debian 12+, Fedora 36+, Arch); the Vulkan + WebGPU one needs 2.38 or newer.
Nothing has to be installed first — not even libfuse2.

# The file you downloaded: -vulkan.AppImage or -vulkan-webgpu.AppImage
APP=./VoxCtrl-linux-x86_64-vulkan.AppImage
chmod +x "$APP"
# Optional: installs the desktop entry and text-injection helpers.
# Global shortcuts need no permissions — your desktop handles them.
"$APP" --install
# Launch the application
"$APP"

What's Changed

• Add Phonon-2 to the Parakeet backend; chunk long audio; stream downloads by @JRufer in #169
• Bump version to 0.8.3 by @JRufer in #170

Full Changelog: v0.8.2...v0.8.3