Introduction

Your voice is the most personal data you’ve got—so why rent it back from a cloud meter? VoiceStudio (previously OmniVoice Studio) flips that script: an open-source desktop studio for zero-shot cloning, voice design, cinematic video dubbing, dictation, transcription, and long-form audiobooks—running on your machine. No account. No API key for the core workflow. Nothing leaving the box unless you opt in.

Latest stable cut as of this scoop: v0.5.2 (September 10, 2026). Active beta energy, AGPL-3.0 app, and a GitHub repo that’s been ripping up Trendshift. If you’ve been craving ElevenLabs-class workflows without the subscription leash, the pocket just opened wide.

What’s Inside the Studio

VoiceStudio isn’t one model bolted into a UI—it’s a multi-engine orchestrator. Switch TTS/ASR from the Model Catalogue (or Ctrl/Cmd+E) and keep the same clone → dub → export flow.

Core workflows:

  • Voice cloning — clean 3-second clip works; 5–15 seconds usually hits harder for the prompt
  • Voice design — describe age, accent, pitch, style, and delivery; spin a new persona without a sample
  • Video dubbing — transcribe → translate → re-voice → export, with speaker casting and timing review
  • Stories & audiobooks — multi-voice scripts, EPUB/PDF import, chapter render, .m4b export
  • Dictation widget — system-wide shortcut, live STT, optional local-LLM cleanup
  • Extras — Demucs vocal isolation, Pyannote/WhisperX diarization, batch queue, AudioSeal watermark tools

Engine matrix (capability-specific): 16 TTS (default OmniVoice / k2-fsa, CosyVoice 3, GPT-SoVITS, VoxCPM2, IndexTTS 2.5, MLX-Audio, and more) and 11 ASR (WhisperX default, Faster-Whisper, MLX Whisper, Parakeet, Moonshine, …). Language catalogue claims 646 TTS languages—real coverage and quality still depend on the engine you load.

Platforms: macOS 13.3+ Apple Silicon, Windows 10/11 x64, Linux x86_64 (glibc 2.39+), plus Docker. Compute: CUDA, Apple Silicon MPS/MLX, Linux ROCm, or CPU. Recommended: 16GB+ RAM, 8GB+ VRAM for the default multi-stage path; CPU mode still works for lighter engines.

Under the hood: Tauri v2 shell + React UI talking to a FastAPI backend on localhost:3900, with OpenAI-compatible audio endpoints and an MCP server for Claude/Cursor agents. Point an OpenAI client at http://localhost:3900/v1 and keep synthesis on-box.

License nuance matters for commercial drops: the app is AGPL-3.0, but downloaded models keep their own terms. Default OmniVoice weights are labeled CC-BY-NC on Hugging Face (tokenizer terms are separate)—review before you ship paid work.

Pro Tip

Clone with one clean speaker, close-mic, no music/reverb in the reference. Longer isn’t always better—zero-shot treats the clip as a prompt, not training data. Match the tone and pace you want in the bounce. For commercial use, check the selected engine’s license before the invoice hits.

How to Get Started with VoiceStudio

  1. Grab the latest release (or site installer) for your OS: GitHub Releases — or voicestudio.sh
  2. Launch, let it create the managed Python env and pull the default model on first run.
  3. Open Voice Cloning, drop a clean 5–15s sample, enter text + language, hit Generate.
  4. Explore dubbing, voice design, and audiobook tabs when you’re ready; flip engines with Ctrl/Cmd+E.

Quick installers also ship from the project site:

curl -fsSL https://voicestudio.sh/install | sh

Docker (linux/amd64; prefer native macOS app on Apple Silicon for GPU):

docker run -d -p 127.0.0.1:3900:3900 -v omnivoice-data:/app/omnivoice_data --name voicestudio palashdeb/omnivoice-studio:stable

OpenAI-compatible local TTS smoke test:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:3900/v1", api_key="local")

with client.audio.speech.with_streaming_response.create(
    model="tts-1",
    voice="<profile-id>",
    input="Made on my own hardware.",
    response_format="wav",
) as response:
    response.stream_to_file("speech.wav")

Repo: github.com/debpalash/VoiceStudio · Site: voicestudio.sh · Release: v0.5.2

Conclusion

VoiceStudio is the local voice suite that finally treats privacy like part of the mix—clone, design, dub, and long-form on your own hardware, with a real engine catalogue instead of one locked cloud stack. Active beta, AGPL app, watch those model licenses—and then let that unfiltered motion hit your monitors without a usage counter.

Open the release, drop a clean sample, and own the pocket.

—Anabel ♡