VoiceStudio: Local Voice Cloning and Dubbing in 646 Languages

VoiceStudio is a desktop application that does voice cloning, voice design, video dubbing, dictation, transcription and audiobook production, and does all of it on hardware you already own. There is no account, no API key, and no usage meter on the local workflow.

It is a Tauri v2 shell wrapped around a React interface and a FastAPI backend, and it has passed 30,000 GitHub stars since the repository was created in April 2026. That is unusually fast for a desktop app, and it reflects a real gap: plenty of open-source speech models exist, but very few of them come with an interface a non-engineer can open and use.

The distinction matters for what follows. VoiceStudio is not a speech model. It is the studio around sixteen of them, plus eleven transcription engines, with a model catalogue, a job queue, and a routing layer that decides whether work lands on CUDA, Apple's MPS and MLX, ROCm, or plain CPU.

If you went looking for OmniVoice Studio and landed somewhere confusing, this is the project. It was renamed to VoiceStudio, and most of the coverage still on the web was written under the old name.

The rename is worth untangling, because three different things share the "OmniVoice" name and they are easy to confuse:

• OmniVoice is a text-to-speech model from k2-fsa, a zero-shot multilingual model covering more than 600 languages. It is a model, not an app.
• OmniVoice Studio was this project's original name. It is now VoiceStudio.
• omnivoice.app is a separate hosted web service, unrelated to this repository.

VoiceStudio ships the k2-fsa OmniVoice model as its default engine, which is where the shared name came from. Traces of the old identity are still visible in the codebase: the Docker image is published as palashdeb/omnivoice-studio, the data directory is omnivoice_data/, and the Colab notebook is still OmniVoice_Studio_Colab.ipynb. None of that is broken, it is just a project that outgrew its first name.

"Local" gets used loosely. In VoiceStudio it means something specific and checkable.

The desktop app talks to a backend bound to loopback only, on localhost:3900. Loopback calls need no server key. Anything beyond loopback, meaning remote workers or an external transcription endpoint, is opt-in, requires a share PIN or API key, must use HTTPS, and will not follow redirects. Analytics is off until you consent, and even when enabled it sends allowlisted usage metadata with no text, audio, file names or projects attached.

In practice that gives you three things a hosted service cannot:

1. Nothing leaves the machine. Recordings, transcripts, cloned voice profiles and projects stay on local disk. For anyone handling client audio, medical dictation or unreleased material, this is the entire argument.
2. No per-character meter. You supply the electricity and the GPU. Generating a twelve-hour audiobook costs the same as generating one sentence.
3. It works offline. Once the model weights are downloaded, the machine can be disconnected.

The honest trade is that you also own the problems. Model downloads are large, first launch builds a managed Python environment, and generation speed is entirely a function of your hardware. A hosted service hides all of that.

646 languages is the headline figure, and it is the one worth being careful about. It is a catalogue size, not a quality guarantee.

The number comes from the default OmniVoice engine, which covers 600 or more languages. The project's own documentation is direct about the caveat: actual coverage and quality depend on the engine you select. The other fifteen engines are nowhere near that range. CosyVoice 3 handles 9 languages plus 18 dialects, VoxCPM2 covers 30, MOSS-TTS-Nano 20, dots.tts 24, Supertonic 3 covers 31, and KittenTTS is English only.

So the accurate reading is this: VoiceStudio can attempt speech in 646 languages through one specific engine, and a long tail of those will be far rougher than the handful of well-resourced languages the models were mostly trained on. If your target language is Urdu, Yoruba or Tagalog, treat the catalogue entry as a starting point and test it before committing to a project.

For the major languages, and for English in particular, the output is strong enough that the comparison to paid services is fair.

Cloning in VoiceStudio is zero-shot, and understanding that one word saves a lot of wasted effort.

Zero-shot means the reference clip is a prompt, not training data. The model is not fine-tuned on your voice. It listens to a short sample and conditions its output on the timbre it hears. Nothing is trained, which is why it finishes in seconds rather than hours.

The practical consequences:

• Three seconds of audio is enough to work. Five to fifteen seconds usually produces a noticeably better result.
• A longer clip does not keep improving the clone. Thirty minutes of audio is not better than fifteen seconds, and can be worse if the extra material is inconsistent.
• Clip quality dominates. One speaker, close to the microphone, no music, no background noise, no reverb.
• Match the delivery you want. The clone copies pace and tone from the prompt, so a flat reference produces a flat read.

There are two other modes alongside cloning. Voice design builds a voice from a written description of age, accent, pitch and delivery style, with no reference audio at all. Convert does speech-to-speech, keeping your performance and timing while changing the voice.

If you are preparing a script to narrate, it is worth running it through a text cleaner first. Text-to-speech engines read stray markup, smart quotes and footnote markers aloud, and those artifacts are far easier to strip before synthesis than to edit out of the audio afterwards. A proofreading pass is cheaper at the script stage for the same reason.

Voice cloning is the demo, but it is not the largest part of the app.

Video dubbing is a full pipeline rather than a single step. VoiceStudio transcribes the source, translates it, keeps speakers separate through diarization, synthesizes the new track while preserving each speaker, and exports the video. Completed dubs flag timing problems for review rather than quietly shipping audio that drifts out of sync. You can import by file or by URL, bring in existing captions, and edit the timed transcript against a zoomable waveform.

Dictation is a system-wide widget on a global shortcut, with live transcription and optional cleanup from a local language model. It inserts text into whatever application has focus.

Audiobooks and stories handle multi-voice scripts with EPUB and PDF import, chapter rendering and .m4b export. A casting board assigns a voice to each speaker. This is the workflow where the absence of a usage meter matters most, since a full book is hundreds of thousands of characters and would be expensive per-character anywhere else. If you are scoping a project, a quick word count gives you a usable runtime estimate, because narration lands somewhere near 150 words per minute.

Vocal isolation via Demucs separates speech from background, and speaker diarization through Pyannote or WhisperX assigns segments to speakers. A batch queue runs large sets of jobs, or watches a folder and picks up new videos automatically.

The engine list is the part that rewards a few minutes of reading, because the defaults are not always right for your hardware. Engines differ in language coverage, whether they can clone at all, which platforms they run on, and crucially their licence.

Engines that cannot clone also cannot preserve a reference speaker during dubbing. VoiceStudio rejects those jobs rather than silently swapping in a different voice, which is the right call and worth knowing before you queue a batch.

The project's own hardware guidance is short and practical:

• Apple Silicon: MLX-Audio or OmniVoice on MPS for text-to-speech, MLX Whisper or Parakeet MLX for transcription.
• NVIDIA with 8 GB VRAM or more: OmniVoice or CosyVoice 3, with WhisperX for word-level timing and diarization.
• Low VRAM or CPU only: PocketTTS, Sherpa-ONNX or KittenTTS, paired with Moonshine or Faster-Whisper in int8.

On the transcription side there are eleven engines, with WhisperX as the default. It is the right default for dubbing and subtitles specifically because it produces word-level timestamps, which is what makes accurate alignment possible.

The minimum is more modest than the feature list suggests, because a GPU is optional.

Platform support is macOS 13.3 and later on Apple Silicon, Windows 10 and 11 on x64, and Linux x86_64 with glibc 2.39 or newer. Docker images are published for Linux AMD64 with CUDA, ROCm and CPU profiles.

Two limitations catch people out. Intel Macs cannot run the local backend at all, because the required PyTorch wheels are not available. They can only connect to a remote backend. And ROCm is Linux only, so AMD and Ryzen AI on Windows falls back to CPU.

Installation is a download from the releases page. First launch creates a managed Python environment and pulls the default model, so expect it to take a while and to use bandwidth. On macOS the first open needs a one-time right-click and Open, because the app is not notarized through the usual path. If you would rather not install anything, a Google Colab notebook runs the whole app on a Colab GPU, though that is remote compute and your audio no longer stays local.

This is the part that makes VoiceStudio interesting to developers rather than only to creators, and it is underplayed in most coverage.

The backend exposes an OpenAI-compatible audio API. Migrating an existing integration is a one-line change:

- base_url="https://api.openai.com/v1"
+ base_url="http://localhost:3900/v1"

Anything already written against OpenAI's speech endpoints keeps working, now against local compute with no per-request cost. The endpoints cover POST /v1/audio/speech for synthesis, POST /v1/audio/transcriptions for transcription, and a WebSocket at /v1/audio/transcriptions/stream for live transcription with partial and final events. Output formats include mp3, opus, aac, flac, wav and pcm.

curl http://localhost:3900/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "tts-1", "input": "Made on my own hardware.", "voice": "default", "response_format": "wav"}' \
  --output speech.wav

There is also an MCP server mounted at http://localhost:3900/mcp, exposing generate_speech, clone_voice and transcribe as tools for Claude Desktop, Cursor and other MCP clients. A stdio shim is bundled for clients that need it. That means an AI agent can narrate or transcribe through local compute without any cloud speech provider in the loop.

If you are building against these endpoints, the request and response bodies are plain JSON, and a JSON formatter makes debugging a malformed verbose_json transcription response considerably faster than reading it as one line.

This is the section worth reading twice, because "free and open source" is doing a lot of work in most write-ups and the reality has an edge to it.

Two separate licences apply, and the permissive one is not the one that governs your output.

The VoiceStudio application is AGPL-3.0. You can run it, modify it and use it internally. If you modify it and offer that modified version as a network service, AGPL obliges you to publish your corresponding source under the same licence. A commercial licence for VoiceStudio's own code is available for proprietary embedding.

The model weights are governed separately, and the application licence does not override them. The default OmniVoice engine ships pretrained weights labelled CC-BY-NC, meaning non-commercial. It also bundles an audio tokenizer under separate Boson Higgs Audio 2 and Meta Llama community terms.

Put plainly: if you generate audio with the default engine and sell it, you are very likely outside the weights licence, even though the app itself places no restriction on selling generated audio.

The way around it is straightforward once you know to look. Several engines ship under genuinely permissive licences: CosyVoice 3, VoxCPM2, MOSS-TTS-v1.5 and dots.tts are all Apache-2.0, and GPT-SoVITS is MIT. If the work is commercial, pick one of those and verify its terms yourself rather than relying on the default.

Two more licence traps in the engine list. IndexTTS 2.5 requires a separate written licence from Bilibili above 100 million monthly active users or roughly 1 billion RMB in annual revenue, which will not affect most readers but is worth knowing. PocketTTS is gated and CC-BY-4.0, and shows its terms before first use.

The repository describes itself as a local ElevenLabs alternative, and that framing is fair as long as the trade is stated honestly in both directions.

Choose VoiceStudio when privacy is non-negotiable, when volume is high enough that per-character pricing hurts, when you need to work offline, or when you want to swap engines and route models yourself.

Stay with a hosted service when you need a specific proprietary voice, when you want guaranteed latency without owning a GPU, when the work is commercial and you would rather not audit model licences, or when you simply do not want to manage a 20 GB model directory.

One caveat on maturity. The README carries an active beta warning and notes an ongoing Electron rewrite, with a request not to file desktop-app issues during it. Use the tagged releases rather than main for anything that matters.

Zero-shot cloning from three seconds of audio is genuinely useful and genuinely open to abuse, and the project does not pretend otherwise.

VoiceStudio embeds AudioSeal watermarking by default. The watermark is imperceptible, survives normal handling, and lets synthetic speech be detected and identified later without changing how the audio sounds. Leaving it enabled costs you nothing.

The consent rule is simple and not optional: clone or synthesize a voice only with the explicit permission of the person whose voice it is. Cloning a real person without consent is, depending on where you live, a route into impersonation, fraud and right-of-publicity claims, and the fact that the tool runs locally does not change any of that.

The project is AGPL-3.0, free, and has no paid tier. It is developed by one maintainer with donations through Ko-fi and PayPal. If it saves you a subscription, that is worth remembering.