VoiceStudio: A Fully Local, Open-Source Voice AI Workbench
Explore VoiceStudio, a local ElevenLabs alternative for voice cloning, dubbing, dictation, transcription, and audiobooks across 646 languages.
VoiceStudio: A Fully Local, Open-Source Voice AI Workbench
Cloud voice platforms are convenient, but they create trade-offs around privacy, recurring costs, network dependency, and control over your data. VoiceStudio takes a different approach: it is an open-source voice workflow engine and desktop application designed to run speech generation and related media tasks on your own hardware.
The project positions itself as a fully local alternative to ElevenLabs, combining voice cloning, voice design, video dubbing, dictation, transcription, audiobook production, and automation in one workspace. It supports workflows across 646 languages, while allowing users to choose from multiple speech engines instead of being locked into a single provider.
What VoiceStudio does
VoiceStudio is broader than a text-to-speech demo. Its workspaces are organized around the complete production workflow:
- Voice cloning: Generate speech from a clean, authorized reference recording.
- Voice design: Describe the voice you want and create a suitable voice without starting from an existing speaker.
- Video dubbing: Produce timed translated speech for video content.
- Dictation: Use a floating widget to turn speech into text.
- Transcription: Process recorded audio locally.
- Stories and audiobooks: Create longer-form narrated content and batch jobs.
- Local API and MCP: Connect voice workflows to scripts, coding agents, and other automation systems.
- Optional remote workers: Offload processing to remote GPU workers when local hardware is not sufficient.
The default setup is powered by k2-fsa/OmniVoice, but VoiceStudio includes an engine catalog so users can select another supported backend. This separation between the workflow layer and the speech engine is important: it lets the application provide a consistent interface while hardware support, quality, speed, and language coverage vary by engine.
Local-first architecture
VoiceStudio’s main design principle is that local workflows run on your own computer. Remote services are optional, and usage analytics require consent. That makes the project useful for developers working with private recordings, internal documentation, unpublished scripts, or customer data that should not be uploaded to a third-party API.
Local processing also changes the economics of repeated generation. Once the models and dependencies are installed, producing additional audio does not require a per-character API charge. The trade-off is that you must provide enough CPU, RAM, storage, and possibly GPU memory for the selected engine. Hardware requirements vary, so the project recommends checking its performance documentation rather than assuming that a GPU will automatically be used.
VoiceStudio can also expose a local backend for programmatic use. The default health endpoint is:
http://localhost:3900/health
After starting the backend, its generated API documentation can be discovered at:
http://localhost:3900/openapi.json
This makes it possible to build reproducible integrations instead of driving the graphical interface manually.
Desktop application and Electron migration
The Electron application is now the primary desktop experience. Release 0.5.3 introduced Electron and was the final Tauri release. Existing Tauri users must install the Electron application separately; the two environments should not be treated as an automatic in-place migration.
The desktop interface includes dedicated screens for voice cloning, dubbing, voice design, and local model management. The model manager is particularly useful for keeping setup separate from production: users can install the required model when prompted, inspect available engines, and reuse downloaded assets across workflows.
The project also provides platform-specific installation guides for:
- macOS
- Windows
- Linux
- Docker
For users who prefer an automated setup, VoiceStudio includes agent-oriented installation guidance. A coding agent can inspect the machine’s operating system, CPU architecture, GPU, RAM, VRAM, free disk space, and existing VoiceStudio installation before choosing a compatible setup. This is a practical approach because speech models can be large, and blindly downloading a model or selecting an unsupported acceleration backend often leads to failed installations.
Running VoiceStudio from source
To launch the Electron preview from source, clone the repository and install its JavaScript dependencies with Bun:
git clone https://github.com/debpalash/VoiceStudio.git
cd VoiceStudio
bun install
bun run dev
The repository combines several technologies. Its implementation is primarily Python, JavaScript, and TypeScript, with Rust used for native components and smaller portions written in CSS and shell scripts. The tree includes separate backend, frontend, Electron, native desktop bridge, deployment, examples, notebooks, tests, and model-related directories.
Before running from source, consult the Electron setup documentation for prerequisites and backend configuration. For a packaged installation, the project recommends downloading the latest stable installer for the target operating system and architecture.
A practical first workflow
A sensible first test is deliberately small:
- Install VoiceStudio using the platform guide.
- Open the voice-cloning workspace.
- Select the bundled demo voice or add an authorized, clean reference recording.
- Enter a short test sentence.
- Install the required model when prompted.
- Confirm which execution device is actually being used.
- Generate an audio file and verify the output path.
Checking the real execution device matters. A machine may have a GPU installed but still run inference on the CPU because of an incompatible runtime, missing acceleration library, or unsupported model. Recording the selected engine, actual device, model location, and generated audio path makes future troubleshooting much easier.
For automation, the repository also publishes installable agent skills:
npx skills add debpalash/VoiceStudio
The voicestudio skill focuses on audio workflows, while voicestudio-maintainer is intended for repository maintenance. If an agent does not support skill installation, the corresponding SKILL.md files can be followed directly.
Privacy, licensing, and responsible use
Voice cloning has obvious consent and identity implications. VoiceStudio is licensed under AGPL-3.0, while individual models may have separate licenses. Review both before using generated audio commercially or distributing a modified version of the software.
Most importantly, clone voices only with permission. A local workflow improves data control, but it does not remove the ethical or legal responsibility associated with impersonation, disclosure, and rights of publicity. Keep reference recordings secure, avoid using voices without authorization, and clearly label synthetic or translated speech where appropriate.
Who should use VoiceStudio?
VoiceStudio is a strong fit for developers and creators who want a self-hosted speech stack rather than a single API call. It is especially relevant for:
- Privacy-sensitive transcription and dictation
- Audiobook and narration pipelines
- Multilingual video dubbing
- Offline or unreliable-network environments
- Local-first AI applications
- Agent-driven audio automation
- Teams that need repeatable, scriptable voice generation
With more than 32,000 GitHub stars, approximately 3,900 forks, 79 contributors, and 39 releases listed on the repository, VoiceStudio has evolved into a substantial open-source project rather than a small proof of concept. Its combination of a desktop interface, local models, API access, MCP support, and optional remote workers gives developers multiple ways to adopt it—from a single manual voice generation to an automated production pipeline.