Repository navigation
Voice Changer 0.3.0
The "runs everywhere, runs smoothly" release.
Added
- Portable builds. The x86_64 release binary is linked against glibc 2.35, so it
starts on Ubuntu 22.04, Debian 12, Linux Mint 21, Zorin 17 and everything newer
(previous releases needed glibc 2.39). A native aarch64 build (Raspberry Pi 5,
Asahi, ARM laptops) ships alongside, each as tarball,.rpmand.deb, plus a
Flatpak bundle and an AUR recipe (packaging/aur). Every release carries a
SHA256SUMSfile. The release workflow refuses a binary that needs a newer glibc or
libstdc++ than that (scripts/check-portable.sh). voice-changer doctorchecks the machine the way a support person would: is
PipeWire the sound server (or only PulseAudio?), is WirePlumber running, can virtual
devices be created, which microphones, GPU and AI runtime are present, are the base
models and voices installed, is there a tray and a shortcut portal. Each problem comes
with the command that fixes it.voice-changer --self-testruns the engine through every preset on synthetic
speech and fails on silence, NaN or runaway levels. CI and packagers use it; the CI
also starts a private PipeWire session, records the virtual microphone through the
PulseAudio API the way Discord does, and checks that the voice came out changed.- AI threads follow your hardware. The number of inference threads is derived from
the processor (physical cores, hybrid P/E cores, SMT) and whether a GPU is used,
keeping one core free for audio and the window; it used to be a fixed 8, too many for
a laptop and too few for a workstation. Settings → AI shows what the machine has and
lets you pick a number, bounded by the threads actually present;--ai-threads N|auto
does the same from the terminal. - Welcome card on the first start: microphone, hear yourself, send it to your apps,
each with its own live checkmark. - The sidebar and the Mixer page show how many apps are hearing the voice, and say
what to do when none is. - An unplugged microphone now falls back to the system default and comes back when
it is plugged in again, with a line in Settings while it is away.
Fixed
- Starting without a working audio system no longer exits silently. When PipeWire is
missing, refused, or PulseAudio still owns the sound, the window opens with a plain
explanation and a Try again button; headless runs print the same text. Every
set-up failure inside the audio thread is now reported, not only the first connect. - A PipeWire or WirePlumber restart no longer leaves a dead app. The audio side
reconnects by itself with a short back-off, re-creates the virtual microphone, and
re-applies the chosen microphone and "Replace my microphone everywhere". The AI
models stay loaded across the gap. - "Replace my microphone everywhere" saved from a previous run now actually applies
at start. It used to race the session manager and silently do nothing; it now waits
for the session manager and a real microphone, then takes over (and never pins the
capture to itself, so no feedback loop). - A missing or broken AI voice no longer makes the worker retry twenty times a
second (and, on GPU builds, leak a model session each time). A failed voice is left
alone for thirty seconds or until you change something; a voice that keeps failing
during inference is parked with a clear message. - Base-model downloads are checked against their size and SHA-256; a cut-off or corrupt
download is deleted and reported instead of being installed as a model that cannot load. - ONNX Runtime libraries are compared by version number (an old
1.9library sorted
after1.22by string order), and libraries older than the API this build needs are
skipped. GPU probing (rocminfo,nvidia-smi) runs in the background, so the window
never waits for it. - The window no longer freezes on the Settings page when a very long log record arrives
(iced logs its whole window configuration, icon pixels included; the in-app log now
keeps one bounded line per record). - A tray that cannot be created no longer cancels the global shortcut, and
"Start minimized" without a tray opens the window instead of leaving no way in. - A crash restores every application's own microphone and the default source before
the process exits. A capture buffer with a bad offset can no longer panic the
realtime thread, and underruns are counted rather than logged from it. - The AI resampler follows the engine sample rate instead of assuming 48 kHz; the
misleading "need contentvec_768l12_q8.onnx" message names the files actually used. - With the window closed to the tray the app ticks at 4 Hz instead of 60 Hz.
Changed
- A second launch brings the running window to the front instead of starting a second
instance. - The monitor ("Hear myself") stream is tagged as music rather than communication so a
Bluetooth headset is not switched to its low-quality profile by a silent stream. - Technical wording in the window was replaced with plain language; the AI runtime line
is now "AI engine".
Downloads
| Fedora, openSUSE, RHEL family | voice-changer-0.3.0-1.x86_64.rpm then sudo dnf install ./voice-changer-0.3.0-1.x86_64.rpm |
| Debian, Ubuntu 22.04+, Mint, Zorin, Pop!_OS | voice-changer_0.3.0-1_amd64.deb then sudo apt install ./voice-changer_0.3.0-1_amd64.deb |
| Any x86_64 Linux with PipeWire | voice-changer-linux-x86_64.tar.gz, unpack and run ./voice-changer |
All three contain the same portable build (needs glibc 2.35 or newer) with the CLAP plugin and the CPU ONNX Runtime fallback; AMD GPUs use the system onnxruntime-rocm, NVIDIA GPUs scripts/get-onnxruntime.sh --cuda. Verify with sha256sum -c SHA256SUMS. Run voice-changer doctor if anything looks off. aarch64 and Flatpak packages will be added once their builds are green.