How to Install VoiceStudio for Local Voice Cloning via Docker
A step-by-step guide to installing VoiceStudio with Docker for local voice cloning, dubbing, and audiobooks, plus how to swap in better TTS engines.
What is VoiceStudio and what does it do?
VoiceStudio is an open-source voice toolkit, released under the AGPL 3 license, that runs entirely on your own hardware instead of calling a cloud API. It bundles voice cloning, voice design, video dubbing, dictation, transcription, and audiobook creation into one interface. The project’s documentation claims support for 646 languages, though actual language coverage depends on which underlying model (engine) you choose. It ships with a local API and MCP (Model Context Protocol) support, so AI agents can call its voice functions directly instead of a human driving the UI. You can run it as a desktop app or deploy it with Docker, which is the faster path if you’re comfortable with a terminal.
By default, VoiceStudio uses OmniVoice from K2FSA as its text-to-speech engine, but the real value of the project is that it’s engine-agnostic. You can swap in other models from its catalog, and quality varies a lot between them.
TL;DR
- VoiceStudio is a self-hosted voice toolkit covering cloning, dubbing, dictation, transcription, and audiobook generation, licensed under AGPL 3 and runnable via Docker or as a desktop app.
- The default engine (OmniVoice) is noticeably weak, producing garbled, low-quality speech, so the out-of-the-box experience undersells what the tool can actually do.
- Swapping engines is where the real quality lives, with models like VoxCPM offering around 30 languages and much more natural voice cloning, though some engines require manual installation inside the container because the catalog’s one-click install doesn’t fetch the weights.
- Hardware matters a lot, text and audio tasks ran fine on a single RTX GPU with 48GB of VRAM, but video dubbing specifically needs roughly 80GB of VRAM for most models.
- Installation requires a few manual steps, including generating a local API token to protect the web UI, running the Docker container with GPU access and persistent volumes for data and model cache, and completing a setup wizard that probes disk space, GPU, and network before you pick starter models.
- It’s a local alternative to cloud TTS services, not a drop-in replacement for something like ElevenLabs, since the bundled default models fall short of commercial-grade voice quality.
- MCP and API support make it agent-friendly, so once it’s running, other automated workflows or tools can trigger voice generation without a human in the UI loop.
How do you install VoiceStudio with Docker?
The Docker route is the quickest way to get VoiceStudio running on a Linux machine with an NVIDIA GPU. The general flow looks like this:
- Generate a local API token. This is a random key that protects the web UI from unauthorized access. For a one-off local test it doesn’t need much ceremony, but if you’re running VoiceStudio in anything resembling production, store that token somewhere safe rather than leaving it in plain text.
- Start the container with GPU access. You need to pass GPU access through to Docker and mount volumes for data and model cache so downloaded models persist across restarts instead of re-downloading every time you spin the container up.
- Wait for the image and initial dependencies to download. First boot takes a few minutes depending on your connection.
- Run a health check. VoiceStudio exposes a health endpoint you can hit to confirm the service came up cleanly before you open the browser.
- Open the web UI and authenticate with your token. The first time you connect, it prompts for the API key you generated earlier.
- Walk through the setup wizard. VoiceStudio probes your system, disk space, GPU, network, and runtime, and flags whether everything checks out before letting you continue.
- Pick a starter TTS model. The catalog presents a range of options. The setup flow also lets you paste in a Hugging Face token if the models you want are gated.
- Choose a transcription model. Options here include Whisper Tiny and other lightweight transcription engines for dictation and captioning features.
Once that wizard finishes, you land on what the project calls the “launchpad,” the main hub for voice cloning, design, dubbing, dictation, and audiobook tools.
Is the default setup good enough to use?
Not really, and this is the most important practical takeaway for anyone installing VoiceStudio. The default engine, OmniVoice, and the default transcription models are described as the weak link of the whole package. In a direct test, cloned audio generated with the default engine came out garbled and difficult to understand, nowhere near natural speech. The tool itself (the interface, the pipeline, the feature set) is solid, but the models VoiceStudio ships with out of the box undersell it badly.
The fix is to go into the model catalog (accessible from the bottom-left of the UI) and swap engines. Catalog entries exist for speech synthesis, voice design, and transcription separately, and you can mix and match. For example, switching the TTS engine to VoxCPM, a model supporting around 30 languages with voice design capability, produced noticeably more natural, intelligible cloned speech than the default.
How do you install a better engine like VoxCPM?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Not every catalog entry installs cleanly from the UI. Some, including VoxCPM in this case, show up in the catalog but fail to install because the one-click flow doesn’t actually pull down the model weights. When that happens, you have two options: follow the project’s own installation guide (which can be sparse and hard to follow), or drop into the terminal and install the package manually inside the already-running Docker container.
The manual path looks like this: with the container running, install the engine package directly inside it, wait for the installation to finish (this can take a couple of minutes), then restart the container. After restarting, go back into the model catalog in the UI, reload, re-enter your API key, and the newly installed engine should now appear as “ready to use” rather than greyed out. From there, selecting it switches your active TTS engine for all downstream tools, cloning, dubbing, audiobooks, without any further configuration.
What can you actually build with it?
Once a decent engine is active, VoiceStudio’s feature set covers most common voice workflows:
- Voice cloning from audio. Upload a sample clip, type or paste a script, and synthesize new audio in that voice. Quality and emotional range depend heavily on whether the selected model supports emotion/style control.
- Voice design and conversion. Build a voice from scratch or convert an existing recording’s characteristics.
- Video and audio dubbing. Drop in a video or audio file to redub it. This is the most hardware-hungry feature, with most dubbing models needing around 80GB of VRAM.
- Story and audiobook creation. A narrator-style editor lets you assign different voices to different characters, write a script with chapter breaks and pauses, and generate a fully narrated audiobook. Scripts follow a simple format: chapter headings, pauses, and a voice name per line, and VoiceStudio stitches the output into one narration.
- Dictation and transcription. Live microphone dictation and file transcription, powered by whichever Whisper-family model you’ve selected.
- A voice gallery. A library of community and built-in voices you can preview and use, with available languages depending on which engine backs each voice.
Is VoiceStudio worth it compared to cloud TTS?
It depends on what you’re optimizing for. If local control, privacy, and running inference on your own GPU matter more than squeezing out the absolute best voice quality, VoiceStudio is a reasonable toolkit, especially once you’ve swapped out the default engine for something stronger. The MCP and local API support also make it genuinely useful for agentic workflows where you want voice generation triggered programmatically rather than through a web UI.
But as a straight quality comparison to a polished commercial service like ElevenLabs, it falls short. The tool’s architecture is flexible, but out-of-the-box quality is weak, getting good results requires knowing which engines in the catalog are actually worth installing, and some of the better ones require manual terminal work to get running at all. Treat it as a flexible local foundation you tune yourself, not a plug-and-play replacement for cloud voice AI.
Frequently Asked Questions
Does VoiceStudio run without a GPU?
The project is designed around GPU acceleration. The demonstrated setup used a single NVIDIA RTX GPU with 48GB of VRAM for text, audio cloning, and audiobook generation. Video dubbing specifically needs considerably more VRAM, around 80GB for most models, so CPU-only or low-VRAM setups will struggle with the heavier features.
Why does the default voice model sound bad?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
VoiceStudio ships with OmniVoice as its default text-to-speech engine, and in testing it produced garbled, unnatural speech. The tool’s architecture supports swapping in stronger engines from its catalog, such as VoxCPM, which produced clearer, more natural cloned audio.
Can I install any model from the catalog with one click?
Not always. Some catalog entries, including VoxCPM in this case, don’t successfully download their weights through the UI’s install button. In those cases, you need to install the model package manually inside the running Docker container via terminal commands, then restart the container and reload the catalog.
Can other AI tools or agents trigger VoiceStudio automatically?
Yes. VoiceStudio includes a local API and MCP support, which lets external agents or automated workflows call its voice generation functions directly instead of requiring manual interaction through the web interface.
Is VoiceStudio a replacement for ElevenLabs?
Based on hands-on testing, no. While VoiceStudio offers a broad local feature set, cloning, dubbing, audiobooks, and transcription, its overall voice quality, even with better engines swapped in, didn’t match a polished commercial service like ElevenLabs. It’s better understood as a flexible, private, self-hosted alternative rather than a direct substitute.
