
VoiceStudio
A local, open-source voice studio for cloning, dubbing, dictation, transcription and audiobooks; AGPL-3.0 app, non-commercial default OmniVoice weightsOpen source
About VoiceStudio
VoiceStudio is an open-source desktop app for voice cloning, voice design, video dubbing, dictation, transcription and audiobook creation running locally, in 646 languages per the repository description. The app is AGPL-3.0-only with a CLA, plus a closed-source Pro build under a paid Commercial License. Weights of OmniVoice, the default model, are CC-BY-NC (non-commercial), and generated audio carries an AudioSeal watermark, on by default. As of 2026-10-26 the latest stable release is v0.5.6, published on GitHub 2026-09-23 16:53 (UTC+8).

What problem it solves
The repository describes VoiceStudio as the open-source, fully-local ElevenLabs alternative: voice cloning, voice design, video dubbing, dictation, transcription and audiobook creation on the local machine with no cloud dependency. The site FAQ says it needs no account and goes online only when the user asks it to: to check for updates or download a model.
Product shape. The LICENSE-NOTICE scope lists the Electron desktop and web app, a native desktop helper and a FastAPI backend. The README says Electron is now the only desktop app and that 0.5.3 was the final Tauri release. Builds cover macOS, Windows, Linux and Docker, with preview apps for iOS and Android.
The default engine is powered by k2-fsa/OmniVoice, and other engines can be selected. The hosted VoiceStudio Cloud has only an early-access sign-up page, which states Nothing is live yet.
The repository was created 2026-04-10 05:40 (UTC+8), and the site footer gives a former name, OmniVoice Studio. Recent releases: v0.5.4 on 2026-09-21 12:14 (UTC+8), v0.5.5 on 2026-09-22 08:42 (UTC+8) and v0.5.6 on 2026-09-23 16:53 (UTC+8), all stable; the prerelease v0.5.7-beta.2 followed on 2026-10-24 07:10 (UTC+8) and mobile-v0.5.7-beta.2 on 2026-10-24 22:03 (UTC+8).
Maintainers: LICENSE-NOTICE names Yupcha Softwares Private Limited as the company that maintains VoiceStudio and carries a copyright line for Palash Debnath and VoiceStudio contributors. The README Team section lists Palash Debnath as Creator and Core Team and Shimul Debnath as Core Team.
Local by default. Explicit at every boundary.
Core capabilities
Workspaces. The README lists voice cloning and voice design, dubbing with timed speech, a floating dictation widget, stories, audiobooks and batch jobs, reusable recipes and watch folders, and a local API and MCP server for agents. The homepage says a clone needs three seconds of reference audio.
Engines and model management. Beyond OmniVoice, other speech engines can be selected, and the Model Manager shows licence information beside each model, listing commercial model use, commercial output use, redistribution and voice consent separately.
AudioSeal watermark. CHANGELOG 0.4.0 says a provenance watermark is applied on every synthetic-audio route through one chokepoint; 0.5.0 adds an Invisible watermark toggle under Settings, Privacy, on by default, available to everyone and affecting only audio generated after a change. Another entry says that when embedding fails, audio passes through unmarked by design and the full traceback is logged. AudioSeal comes from facebookresearch, with code and model weights under MIT.
Hardware. The README lists CUDA on NVIDIA GPUs, Metal on Apple Silicon and slower CPU-only operation without a dedicated GPU; Windows on ARM is experimental, and an Intel Mac runs only the UI against a backend on another machine. The site FAQ says about 8 GB of RAM is enough to start.
Local first. The privacy page says local recordings, generated audio, transcripts and derived voice data are written to user-controlled storage and never reach the website or its control plane.
Who uses it
As of 2026-10-26 the GitHub API shows 86,585 stars, 9,591 forks and 194 open issues and pull requests combined.
Issue 2586, the CLA sign-up thread, opened 2026-10-03 07:12 (UTC+8) and had 53 comments as of 2026-10-26; it remains open.
Many CHANGELOG entries credit outside contributors by name, across dubbing, cloning, API compatibility and platform setup. The README says contributors sign the CLA before a first pull request.
How it compares
Against ElevenLabs: VoiceStudio calls itself a fully-local ElevenLabs alternative in its repository description, with the desktop app running on the local machine; the framing is VoiceStudio's own.
Against Voicebox: Voicebox is MIT-licensed and describes itself as The open-source AI voice studio. Clone, dictate, create., a similar position. VoiceStudio's app is AGPL-3.0-only, contributions go through a CLA, and a closed-source Pro build exists alongside it.
Against Coqui TTS: Coqui TTS is MPL-2.0, a deep learning toolkit for text-to-speech rather than a desktop app, and GitHub shows its last push on 2024-08-16.
Against OmniVoice itself: the k2-fsa/OmniVoice code is Apache-2.0 and is described as High-Quality Voice Cloning TTS for 600+ Languages; VoiceStudio adds dubbing, dictation, audiobook, local API and model-management workflows on top.
Who it's for / Not for
Good fit:
- Creators and teams that need cloning, dubbing and transcription on their own machines without uploading audio to a cloud service.
- Developers wiring speech capabilities into agents through the local API or MCP.
- Individuals working mainly on non-commercial projects who accept AGPL-3.0 obligations.
Not a fit:
- Commercial output from the default model: OmniVoice pretrained weights are CC-BY-NC, and LICENSE-NOTICE says a VoiceStudio Commercial License does not replace model terms.
- Buyers expecting a list of models cleared for commercial use: no record in model_licenses.json is marked as commercially cleared.
- Closed-source distribution or private hosting of a modified version: the Pro page says neither is part of Pro or Lifetime and both go through a separate enquiry.
- Intel Mac owners with no second machine to run the backend.
Getting started
Setup. The README gives a one-line install script for macOS and Linux and a PowerShell equivalent for Windows; GitHub Releases also offers a dmg for macOS, an exe for Windows and an AppImage or deb for Linux, and Docker is supported. The installer keeps existing settings, projects and models.
First run. The README flow is to open Voice cloning, select a voice or add a clean reference recording, enter text and generate, adding the required model when prompted.
Development and integration. A source checkout runs with bun install, bun run setup:api and bun run dev; docs for the local API and MCP live in the docs folder. The README also carries a prompt intended for coding agents.
Commercial work. The Pro page sells Pro and Lifetime plans, and Pro runs without an account.
Pricing & license
Open-source app: free, AGPL-3.0-only. LICENSE-NOTICE says commercial use under the AGPL-3.0 is free; what is paid is a Commercial License for organizations embedding VoiceStudio in a closed-source or proprietary product or service without AGPL-3.0 copyleft obligations, plus the Pro features.
Pro: $99 per user per year. Lifetime: $299 per user, one-time. Both are the closed-source Pro build with a Commercial License, covering up to 3 registered devices per user and 3 running at once. The Pro page lists multiple devices, connections to other devices and early access, with further items marked Upcoming.
Refunds: the terms page offers a full refund within 7 days of purchase; after that the annual plan can be cancelled before renewal with no partial refund. Payments run through Razorpay, Zoho Payments or PayU.
Enterprise: custom terms, and an enquiry is not a purchase or licence grant. Closed-source distribution and private hosting are outside Pro and Lifetime.
CLA: contributors license contributions to Yupcha Softwares Private Limited. CLA 1.0 lets the maintainer license them under any terms, including the AGPL, other open-source licences and commercial or proprietary licences; it commits to keeping contributions in the public repository under AGPL-3.0 or another OSI-approved licence and does not require publishing source for commercial or proprietary editions.
Model weights: LICENSE-NOTICE says downloaded weights are not relicensed by VoiceStudio and keep their own terms.
Risks & limits
Non-commercial models. model_licenses.json and model-inventory.md mark 4 records non-commercial, and their model cards read:
- k2-fsa/OmniVoice, the default engine named in the README: code under Apache 2.0, the pretrained model under CC-BY-NC because of training data such as Emilia, version not stated.
- Serveurperso/OmniVoice-GGUF, a record in the model catalogue: cc-by-nc-4.0.
- facebook/nllb-200-distilled-600M, a record in the model catalogue: cc-by-nc-4.0.
- mlx-community/Llama-OuteTTS-1.0-1B-4bit, a record in the model catalogue: cc-by-nc-sa-4.0; the upstream OuteAI card places the initial Llama 3.2 components under the Llama 3.2 Community License.
LICENSE-NOTICE adds that the OmniVoice audio_tokenizer carries separate Boson Higgs Audio 2 and Meta Llama community terms and that a VoiceStudio Commercial License replaces none of them.
Model licence clearance incomplete. model-inventory.md says Every production record has missing exact-file and/or terms evidence, and no record claims commercial clearance.
Bundled binaries. LICENSE-NOTICE says the locked PyAV 15.1.0 wheels bundle FFmpeg and x264 and x265 libraries with unresolved redistribution requirements.
Licensing model. The CLA allows contributions to be used in closed-source commercial editions, and the CHANGELOG says compute devices on other machines, workers and GPU sharing moved from the open-source app to VoiceStudio Pro.
Stable cadence. As of 2026-10-26 only prereleases have followed v0.5.6. Cloud has not launched, and the terms page says the hosted service may launch differently or not launch.
Differing statements, listed side by side:
- Commercial licence: LICENSE-NOTICE says what is paid is a commercial license for organizations that want to embed VoiceStudio in a closed-source or proprietary product or service without the AGPL-3.0 copyleft obligations, and the VoiceStudio Pro features; the Pro page sells Pro and Lifetime with a Commercial License and states Closed-source distribution and private hosting aren't in Pro or Lifetime.
- Network use: the site FAQ says the app goes online only when the user asks it to, to check for updates or download a model; the privacy page adds explicitly configured network-backed adapters; CHANGELOG 0.5.7 says a render that cannot fetch the AudioSeal watermark checkpoint retries the load after 10 minutes.
- Payments and law: the terms page says governing law is to be published before payments go live and that counsel sign-off comes before payment is enabled; the same page calls Pro and Lifetime real purchases, and the Pro page offers $99 and $299 checkouts.
Model licence records: the repository file model-inventory.md says model_licenses.json records 63 configured model repositories and seven unresolved asset families (as of 2026-10-26); the 4 non-commercial records are for OmniVoice, OmniVoice-GGUF, NLLB and Llama-OuteTTS; the same file says false does not mean every listed model forbids commercial use.
Our verdict
Inference: VoiceStudio suits people who need cloning, dubbing and transcription on their own machines; the feature set is broad, the local API and MCP make agent integration direct, and the AGPL-3.0 app itself is free for commercial use.
Inference: the commercial limits come mainly from models rather than the app. The default OmniVoice weights are CC-BY-NC, no record in the model licence inventory is cleared for commercial use, and the Commercial License does not cover model terms; the project is open core, contributions can flow into the closed-source Pro build through the CLA, and generated audio carries an AudioSeal watermark.
Sources
Official
- Free Open-Source AI Voice Cloning & Dubbing | VoiceStudio · Accessed Oct 26, 2026
- VoiceStudio Pro: Commercial License for Client Work | VoiceStudio · Accessed Oct 26, 2026
- Hosted Voice and Dubbing API, Early Access | VoiceStudio · Accessed Oct 26, 2026
- Website terms and current product availability | VoiceStudio · Accessed Oct 26, 2026
- Privacy policy | VoiceStudio · Accessed Oct 26, 2026
- VoiceStudio README · Accessed Oct 26, 2026
- GNU Affero General Public License Version 3 · Accessed Oct 26, 2026
- VoiceStudio License Notice · Accessed Oct 26, 2026
- VoiceStudio Individual Contributor License Agreement 1.0 · Accessed Oct 26, 2026
- Model licence inventory · Accessed Oct 26, 2026
- VoiceStudio model licence records (model_licenses.json) · Accessed Oct 26, 2026
- VoiceStudio Changelog · Accessed Oct 26, 2026
- VoiceStudio repository on GitHub · Accessed Oct 26, 2026
- VoiceStudio releases on GitHub · Accessed Oct 26, 2026
- Contributor License Agreement: sign here · Accessed Oct 26, 2026
Media & third party
- k2-fsa/OmniVoice model card on Hugging Face · Accessed Oct 26, 2026
- Serveurperso/OmniVoice-GGUF model card on Hugging Face · Accessed Oct 26, 2026
- facebook/nllb-200-distilled-600M model card on Hugging Face · Accessed Oct 26, 2026
- mlx-community/Llama-OuteTTS-1.0-1B-4bit model card on Hugging Face · Accessed Oct 26, 2026
- OuteAI/Llama-OuteTTS-1.0-1B model card on Hugging Face · Accessed Oct 26, 2026
Related products
- k2-fsa/OmniVoice on GitHub (Apache-2.0) · Accessed Oct 26, 2026
- facebookresearch/audioseal on GitHub (MIT) · Accessed Oct 26, 2026
- jamiepine/voicebox on GitHub (MIT) · Accessed Oct 26, 2026
- coqui-ai/TTS on GitHub (MPL-2.0) · Accessed Oct 26, 2026