OmniVoice is a model you call from Python. VoiceStudio is the application built around it — and around whichever other engine you pick. Same local-only premise, but with the workflows attached: clone a voice, dub a video with timed speech, dictate into whatever window has focus, transcribe, batch a whole audiobook.

It bills itself as the open-source ElevenLabs alternative, and the comparison is fair on capability. AGPL-3.0, Electron desktop app, 646 languages, nothing leaving the machine unless you opt into a remote worker.

Install 

  curl -fsSL https://voicestudio.sh/install | sh          # latest release
curl -fsSL https://voicestudio.sh/install | sh -s -- --version X.Y.Z
curl -fsSL https://voicestudio.sh/install | sh -s -- --uninstall   # keeps your data

macOS and Linux; Windows and Docker have their own guides, and there are plain release downloads. The installer preserves settings, projects, and models across upgrades.

Then: open Voice cloning, pick a bundled voice or drop in a clean reference recording, type your text, generate. It prompts for the model download when it needs one.

There’s also an agent path, which is a neat touch — hand your coding agent the agent guide and it handles hardware detection, reuses existing data, and asks before pulling models:

  Install the VoiceStudio Electron app on this device and verify it works, following
https://github.com/debpalash/VoiceStudio/blob/main/docs/install/agent.md

What it actually does

WorkspaceFor
Voice cloningA voice from a few seconds of reference audio
Voice designDescribe the voice you want in words, no reference needed
Video dubbingSpeech timed to an existing video track
DictationA floating widget — talk, words land at the cursor
TranscriptionThe other direction, locally
Stories & audiobooksLong-form batch jobs rather than one clip at a time
ModelsInstall, switch, and manage engines from the UI

The default engine is OmniVoice (k2-fsa), with others selectable per job — so the model page and this page are the same stack at two altitudes.

Local API and MCP

The part that makes it more than a nice UI:

  • A local HTTP API, so your own scripts can generate speech without a cloud key.
  • An MCP server, so an agent can produce audio as a step in a larger job — narrate a generated video, read back a draft, dub a clip — without you in the loop.
  • Agent skills via npx skills add debpalash/VoiceStudio (voicestudio for audio work, voicestudio-maintainer for repo maintenance).

That combination is the argument for it over a hosted service: the per-character meter disappears, and “generate 400 lines of narration” stops being a budgeting decision.

Where it sits

ElevenLabsVoiceStudio
RunsHostedYour machine
CostPer characterElectricity
CeilingBest-in-class quality, quicklyDepends on your GPU and chosen engine
DataLeaves your machineDoesn’t
SetupAn API keyAn install and a model download

Reach for ElevenLabs when quality is the whole point and volume is low. Reach for VoiceStudio when volume is high, the audio is sensitive, you’re offline, or you want an agent to make audio unattended.

Before you rely on it

  • AGPL-3.0, and — stated separately in the README — the models carry their own licences. Check both before anything commercial; an Apache-2.0 model inside an AGPL app is still two sets of terms.
  • Hardware decides the experience. Requirements vary by engine; the project publishes benchmarks and a performance guide. Read them before planning a big batch.
  • Tauri is gone. 0.5.3 was the last Tauri release; Electron is now the only desktop app and web UI, and old installs need a separate migration.
  • Clone voices only with permission. The project says this plainly and so should anyone using it. A convincing clone of someone who didn’t agree is a problem no licence fixes.

Next

The other direction — turning what you say into text → OpenWhispr

Edit this page

Last updated 27 Sep 2026, 00:00 UTC. history