VoiceStudio 0.5.7: phones, CPUs and a calmer app

The app now fits more machines, and asks less of you

VoiceStudio 0.5.7 is a wide release, and it took longer to arrive than it should have. The release notes open with an apology for the wait and thanks for the patience and the reports that came in meanwhile. They also welcome Shimul Debnath (velixio) to the core team, helping review pull requests, triage issues and keep releases moving.

Three changes matter most if you use the app:

• Phones. Preview iOS and Android apps run models on the phone itself.
• No GPU. PCs without a dedicated GPU get a small CPU build of PyTorch and can run the app.
• Less clutter. The sidebar, Models, Dubbing, Stories and Workflows pages were rebuilt around fewer, clearer controls.

There is also an opt-in Beta update channel, a bottom player for takes, and a GPU acceleration panel that reports Radeon cards honestly. The notes count hundreds of fixes to cloning, dubbing, audiobooks, downloads and dictation; this post does not list them.

VoiceStudio on a phone

The phone apps run speech, dictation, voice cloning and dubbing on the device. Once the models are downloaded they work offline. Pairing a computer is optional.

They are previews. Android has an APK you can install directly, plus an App Bundle for Google Play. On iPhone you get an unsigned IPA that you sign with your own Apple ID, for example through Sideloadly. It is not a TestFlight or App Store build; the notes say App Store signing is blocked by Apple's certificate limit.

No GPU, still welcome

A machine without an NVIDIA GPU now installs the small CPU PyTorch build. It needs about 5 GB free instead of 9 GB, and it loads voice models in float32 instead of float16. Small models install with as little as 2 GB of spare disk; the setup check only warns below 10 GB rather than blocking you.

Windows on ARM gets a native installer, marked experimental. The AI backend runs as x64 under emulation, on the CPU.

If you do have a GPU, Settings → Performance → GPU acceleration lists your cards, the installed PyTorch build, and which engines use the GPU or run on the CPU and why. Setup now names a Radeon card as a Radeon card instead of blaming a missing NVIDIA driver.

A calmer app

Most of the visible work is subtraction. The sidebar is a slim icon rail beside each section's own pane. Settings → Models shows one list per family, so you install, accept the licence for and select an engine on the same row. Recent takes and project audio play in a bottom player with a seekable waveform.

Dubbing has a one-row source, a preview of the loaded file, and files without sound reported as a note instead of an error. Workflows opens templates from a picker, adds steps from a toolbar on the canvas, and flags steps with missing settings before you run. Failed generations and failed engine starts now name a likely cause and lead with a fix.

Smaller additions you may want right away:

• Find and replace in the Dubbing transcript, with the edit undone as one step
• A Background volume slider on dub exports, from 0 to 400%
• Smart Fit speed limits, set from the Timing section
• Portuguese (Brazil) and Portuguese (Portugal) as separate language choices
• Uyghur as a dub target, with right-to-left scripts now reading right to left

The full list, with contributor credits, is in the v0.5.7 release notes.

Before you update

A few changes can stop something that worked yesterday.

• Model licences. Models without a plainly commercial licence, including the default OmniVoice model, ask you to accept their licence once in Model Catalogue. Until you do, API and MCP requests that use them return a licence error.
• Remote compute moved to Pro. Remote compute devices, remote workers and GPU sharing are no longer in the open-source app. Saved projects, voices and settings are unaffected.
• Host names. Reaching Docker, a LAN or a reverse proxy by host name, without an API key or session, needs that name listed in OMNIVOICE_ALLOWED_HOSTS. IP addresses, localhost, *.ts.net and host.docker.internal work as before.
• Private URL imports. Importing from a server on your own network needs OMNIVOICE_ALLOW_PRIVATE_URL_IMPORTS=1.
• Coming from Tauri. Tauri v0.5.3 is the last Tauri release and its updater cannot install Electron. Close the app, back up its data directory, install Electron, and check your voices and projects before removing Tauri. The migration guide has the steps.

Both environment variables go in ~/.config/omnivoice/env on macOS and Linux, or %USERPROFILE%\.config\omnivoice\env on Windows. For Docker, pass them with -e. The API auth guide covers the details.

How to update

Install over your existing Electron app. Voices, projects and settings are kept. If setup asks, choose Install local runtime to refresh its dependencies. Installers are on the download page, and Android and iPhone files are in the release assets. Check downloads against SHA256SUMS.txt.

On trust: the macOS installers are signed and notarized, and the Android build is signed with the VoiceStudio release key. The Windows installers are not code-signed yet, so SmartScreen may warn. Choose More info, then Run anyway. Linux packages are not code-signed.

If you would rather get fixes early, Settings → Updates now has a Stable or Beta choice. Betas are -beta.N prereleases that never reach Stable users, and you roll onto the next stable release automatically. Docker gets matching :beta tags.

First thing to try: open the GPU acceleration panel and read what it says about your machine. If you want to make a voice next, start with the local cloning guide.

Sources and version notes

Written from the VoiceStudio v0.5.7 release notes, published 26 October 2026. Nothing here was measured separately: the 5 GB and 9 GB figures, the 2 GB and 10 GB disk thresholds and the 0 to 400% range are quoted from those notes. The phone apps and the Windows ARM64 installer are described there as preview and experimental, and this post has not tested them.

• v0.5.7 release notes
• Phone apps guide
• Update channels guide
• Tauri to Electron migration guide

The app license does not grant rights to any model or to generated audio; each model's own terms apply.