VoiceStudio Setup: Clone Voices Locally, No Cloud

VoiceStudio is a free, open source desktop app that clones voices, dubs video, transcribes speech, and produces long form audiobooks entirely on your own hardware, no ElevenLabs subscription and no per-character meter required. It just crossed 14,500 GitHub stars and sits near the top of today's trending page, formerly known as OmniVoice-Studio before its recent rename.

• VoiceStudio installs with one command on Windows, macOS (Apple Silicon), or Linux, plus a Docker path for servers; it bundles 16 TTS engines and 11 ASR engines covering a 646-language catalogue.
• Everything runs locally by default: voice cloning, video dubbing, dictation, and audiobook generation never leave your machine unless you explicitly enable a remote worker.
• It exposes an OpenAI-compatible /v1/audio/speech API on localhost:3900, so existing ElevenLabs-style client code can point at your own machine instead of a paid endpoint.
• It is AGPL-3.0 and free, but the default OmniVoice model weights carry a separate CC-BY-NC license, so check the terms of whichever engine you pick before selling generated audio commercially.

The exact steps, start to finish

1. Step 1. Check your machine is ready.

   # Windows PowerShell: confirm winget, OS version, and free disk space
   winget --version
   [System.Environment]::OSVersion.Version
   Get-PSDrive C | Select-Object Used,Free

   You need Windows 10 (21H2 or newer) or Windows 11 x64, roughly 10 GB free disk for the app and its model weights, and winget available since the installer below uses it. An NVIDIA GPU is optional; VoiceStudio falls back to CPU automatically.
2. Step 2. Run the one-line installer.

   $ irm https://voicestudio.sh/install | iex

   This single PowerShell command installs Git, FFmpeg, uv, and Bun through winget, then clones the repository and builds the desktop app for you. No manual toolchain setup needed.
3. Step 3. Launch VoiceStudio and let first boot finish. The first launch creates a managed Python environment through uv, syncs dependencies, and downloads the default OmniVoice model; the splash screen shows live progress for every stage, so let it run instead of force-closing it.
4. Step 4. Add a free Hugging Face token. Create one at huggingface.co/settings/tokens (the "read" scope is enough), then either paste it into Settings → API Keys inside the app, or set it as a PowerShell environment variable:

   [Environment]::SetEnvironmentVariable("HF_TOKEN","hf_yourtokenhere","User")

   The core cloning workflow runs without a token, but the gated diarization model and a couple of the larger TTS engines return a 401 until one is set.
5. Step 5. Clone your first voice. Open Voice Cloning in the app, drop in a clean 5 to 15 second sample of one speaker with no background music or noise, type the text you want spoken, pick a language, then click Generate. That is a fully local voice clone: no upload, no queue, no bill.

What is VoiceStudio and why is it trending?

VoiceStudio, renamed from OmniVoice-Studio earlier this year, is a desktop application built on a Tauri v2 shell in Rust, a React and Vite frontend, and a FastAPI backend that talks to a registry of TTS and ASR engines over localhost. It handles zero-shot voice cloning from a short reference clip, voice design from text instructions like age and accent, full video dubbing with speaker preservation, a system-wide dictation widget, vocal isolation, speaker diarization, and multi-voice audiobook rendering with EPUB and PDF import. The repository sits at more than 14,500 stars, shipped v0.5.1 on August 28, 2026, and is licensed AGPL-3.0. The trending spike lines up with a real gap: ElevenLabs-class voice cloning has mostly lived behind subscriptions and character meters, and VoiceStudio ships it as a native app with GPU auto-detect across CUDA, Apple Silicon MPS and MLX, ROCm on Linux, and CPU fallback everywhere else.

RelatedUnsloth Setup: Run and Train LLMs on Your Own Machine

How do you install VoiceStudio on Windows, macOS, and Linux?

Every platform gets the same one-line installer pattern; only the shell changes. On Windows, run the checklist's PowerShell one-liner above, or download the MSI from the Releases page and choose the per-machine or current-user build depending on whether you have admin rights.

On macOS (Apple Silicon only; Intel Macs can run the UI but not the local backend, since PyTorch dropped Intel-Mac wheels), the equivalent one-liner is:

# macOS, one-liner install from source
curl -fsSL https://voicestudio.sh/install | sh

Or download the DMG that matches your chip, drag VoiceStudio.app into Applications, then right-click it and choose Open once to clear Gatekeeper's quarantine flag on this ad-hoc-signed build. If Terminal is easier, the same fix is one command: xattr -dr com.apple.quarantine "/Applications/VoiceStudio.app".

On Linux (Debian, Ubuntu, Fedora, Arch, or WSL), the installer is identical to macOS:

# Linux, one-liner install from source
curl -fsSL https://voicestudio.sh/install | sh

Or grab the AppImage from Releases if you would rather skip building; it needs glibc 2.39+ and an X11 or Wayland session. AMD GPU owners only get real acceleration on Linux through ROCm, since Windows ships no ROCm build. A Docker image with CUDA, ROCm, CPU, and worker-only profiles also exists for headless or server installs, documented in the project's docs/install/docker.md.

How does VoiceStudio compare with ElevenLabs?

The comparison is deliberate: VoiceStudio's own README bills itself as the fully-local ElevenLabs alternative, and the tradeoff really is that clean.

RelatedOpenCut Setup: Self-Host the Free CapCut Alternative

ElevenLabs still wins on zero setup time and consistently polished voices out of the box. VoiceStudio wins everywhere local control, offline use, or per-minute cost actually matters, which describes long-form audiobook production, dubbing large video libraries, and anyone who would rather own the hardware than rent the API.

What are the gotchas before you rely on it?

Four things worth knowing before you build a workflow around it. First, the README flags VoiceStudio as active beta: the main branch moves fast, so stick to tagged releases for anything you cannot afford to break. Second, AGPL-3.0 is a copyleft license: you can run and modify VoiceStudio freely, but if you modify it and offer that modified version as a network service, the license requires you to publish the corresponding source. Third, the default OmniVoice model weights are CC-BY-NC, separate from the AGPL-3.0 app license, so review the specific engine's terms before using generated audio commercially. Fourth, first launch downloads real weight, budget several gigabytes and a stable connection, and expect the initial model pull to take longer than the install itself on a slow line.

Our take

VoiceStudio's real achievement is not any single engine, since most of the 16 TTS options and 11 ASR options are borrowed from other open source projects like WhisperX, GPT-SoVITS, and Sherpa-ONNX. The achievement is the packaging: one installer, one desktop shell, an OpenAI-compatible endpoint on localhost, and GPU auto-detection that just works across CUDA, Apple Silicon, and ROCm. That is the same trick Ollama pulled for local language models, applied to voice. The AGPL-3.0 license and the CC-BY-NC default model weights are worth reading closely before this becomes part of a commercial pipeline, but for personal use, dubbing, dictation, or audiobook production, this is the easiest private voice stack available today, and the price argument, free versus a metered subscription, is not close.

Original analysis by GenZTech. Tool documentation: debpalash/VoiceStudio on GitHub.