What is VoiceStudio?
VoiceStudio is a desktop app, with Docker images and preview phone apps, that does locally what ElevenLabs does in the cloud: clone a voice from a reference recording, design one from a text description, dub video with timed speech, dictate into any text field through a floating widget, transcribe audio and narrate audiobooks. The README claims 646 languages. The default engine is k2-fsa's OmniVoice, a 600-plus-language text-to-speech model built on Qwen3-0.6B, and a Model Manager offers dozens of other engines.
It runs on Apple Silicon through Metal, on NVIDIA GPUs through CUDA on Windows and Linux, and on the CPU everywhere, more slowly. For agents, the backend mounts an MCP server at /mcp on port 3900 with tools such as generate_speech, clone_voice, design_voice and transcribe, alongside a local REST API that includes an OpenAI-compatible transcription endpoint.
The app is AGPL-3.0 and maintained by Yupcha Softwares Private Limited, which also sells a closed-source Pro build with a commercial licence and extra batch features.
How much does VoiceStudio cost?
| Plan | Price | What you get |
|---|---|---|
| Open source | Free | AGPL-3.0; cloning, design, dubbing, dictation, local API and MCP |
| Pro | $99/user/year | Commercial licence, 3 devices, batch dubbing, watch folders, remote GPU workers |
| Lifetime | $299 one-time | Pro features and future updates, 3 devices |
| Enterprise | Contact sales | Custom terms, such as embedding in closed-source products |
Open-source app free under AGPL-3.0. Pro $99 per user per year or Lifetime $299 one-time, each for one user on up to 3 devices, with a commercial licence and batch features; Enterprise on request. Model weights carry their own licences.
Is VoiceStudio worth it?
VoiceStudio gives one person on one machine most of what a cloud voice service does, and the fine print decides whether you can sell the result. The app code is AGPL-3.0, but its own licence notice says the default OmniVoice weights are CC-BY-NC, and that its model inventory asserts no commercial clearance for any model. The repository has about 92,300 GitHub stars and 10,200 forks since it was created in April 2026, and shipped v0.5.7 on 26 October 2026.
The agent integration is careful in places and open in others. MCP tools can return audio as files rather than base64 so a WAV does not flood the agent's context, and file reads are confined to a base directory set with OMNIVOICE_MCP_BASE_PATH. But the docs state that the backend is loopback-only and unauthenticated by default, that the MCP transport is not authenticated, and that the settings routes are remote-code-execution class and reachable from loopback, so any process on the machine can drive them. LAN sharing adds a PIN that the docs warn travels unencrypted over plain HTTP. The first generate_speech call downloads about 2.3 GB of model if setup has not already done it.
The one-line install, `curl -fsSL https://voicestudio.sh/install | sh`, deserves a pause. On 28 October 2026 that URL redirected to the install script on the main branch, although the installer guide says the site serves a bundled, tested copy rather than mutable main. The script checks the download against the release's SHA256SUMS.txt, which catches corruption but not a compromised release, since both come from the same place. On macOS it also clears the quarantine flag and re-signs the app ad hoc when signature verification fails; an issue asking for signed binaries (#1779) has been open since September 2026. Downloading the script and running `sh scripts/install.sh` after reading it is the documented alternative.
The AGPL itself is permissive for users: commercial and internal use of the app is free, and the copyleft obligation bites only if you modify VoiceStudio and offer it to others over a network. Embedding it in a closed product needs a commercial licence, and contributors sign a CLA assigning rights to the company. Pro costs $99 per user a year or $299 once, and it is where batch dubbing, watch folders and remote GPU workers live, although the README lists watch folders among what the app does. Against audio.cpp, which is a runtime rather than an app, VoiceStudio is the friendlier way to use these models; the licence that matters for paid work is the one on the weights.
Where it earns its place
An independent podcaster on a 16 GB Apple Silicon MacBook
Produce narration in my own cloned voice without uploading recordings to a cloud service
Install the .dmg, add a clean reference clip with its transcript and generate on Metal; the app unloads the TTS model to make room for transcription on 16 GB machines. Free, but a sponsored or paid show should pick an engine whose weights allow commercial output, since the default OmniVoice weights are CC-BY-NC.
A developer giving a Claude Code or Codex agent spoken output
Let the agent read summaries aloud and transcribe voice notes locally
Point the client at http://127.0.0.1:3900/mcp/ or the stdio shim, set OMNIVOICE_MCP_OUTPUT_MODE=files and a base path so audio stays out of context, and bind a voice per agent. No fees; keep it off shared machines because the loopback API has no authentication.
A five-editor localisation studio dubbing training videos
Queue videos and render several target languages without per-minute cloud fees
Batch dubbing needs Pro: $495 a year for five seats, or $1,495 once on Lifetime. The docs say an NVIDIA card with 16 GB or more runs 3 to 4 generations at once, while 10 GB or less runs one at a time. Commercial output still depends on each model's licence.
Before you commit
- Setup
- `curl -fsSL https://voicestudio.sh/install | sh` on macOS or Linux, `irm https://voicestudio.sh/install.ps1 | iex` on Windows, or a .dmg, .exe, .AppImage, .deb, snap or GHCR Docker image. The app asks before downloading any model; OmniVoice is about 2.3 GB.
- Integrations
- MCP over Streamable HTTP at :3900/mcp/ or a stdio shim, with tools generate_speech, clone_voice, describe_voice, design_voice, transcribe and list_voices; OpenAI-compatible POST /v1/audio/transcriptions and a streaming WebSocket; agent skills via npx skills add debpalash/VoiceStudio.
- Data and privacy
- Processing is local and the docs say there are no accounts or updater telemetry. The backend and MCP endpoint are unauthenticated on loopback; LAN sharing uses a PIN sent over plain HTTP unless you add a TLS proxy. App licence AGPL-3.0; default OmniVoice weights CC-BY-NC.
- Cost at scale
- Free to self-run on your own GPU or CPU; CPU is about half the speed of Apple Silicon. Pro is $99 per user a year and Lifetime $299, each for up to 3 devices; closed-source embedding is priced on request.
- Limits
- 40 open issues and pull requests, including multi-speaker dubs merging speakers (#3419) and static noise in generations (#2818); Intel Macs need a remote backend; phone apps are previews; releases up to 0.5.3 used Tauri and need a separate migration.
What's good
- Cloning, design, dubbing, dictation and transcription in one local app
- Runs on Apple Silicon, NVIDIA or CPU, plus Docker
- MCP server and OpenAI-compatible transcription endpoint
- Per-model licence notices in the Model Manager
What's not
- Default OmniVoice weights are CC-BY-NC
- Local API and MCP are unauthenticated on loopback
- curl | sh installer strips macOS quarantine; binaries unsigned
- Batch dubbing and watch folders are Pro-only
Best for
Individuals and small teams who want voice cloning, dubbing and transcription on their own hardware and can check each model's licence before commercial use.
VoiceStudio alternatives
VoiceStudio: common questions
- Is VoiceStudio worth it?
- VoiceStudio is a serious local stand-in for ElevenLabs, but the default voice model is licensed for non-commercial use and the installer asks a lot of trust. We rate it 3.8 out of 5.
- How much does VoiceStudio cost?
- Open-source app free under AGPL-3.0. Pro $99 per user per year or Lifetime $299 one-time, each for one user on up to 3 devices, with a commercial licence and batch features; Enterprise on request. Model weights carry their own licences.
- What are the downsides of VoiceStudio?
- Default OmniVoice weights are CC-BY-NC; local API and MCP are unauthenticated on loopback; curl | sh installer strips macOS quarantine; binaries unsigned; batch dubbing and watch folders are Pro-only.
- Who is VoiceStudio best for?
- Individuals and small teams who want voice cloning, dubbing and transcription on their own hardware and can check each model's licence before commercial use.
- How much does VoiceStudio cost for a team or in production?
- Free to self-run on your own GPU or CPU; CPU is about half the speed of Apple Silicon. Pro is $99 per user a year and Lifetime $299, each for up to 3 devices; closed-source embedding is priced on request.
- Is VoiceStudio safe for company data?
- Processing is local and the docs say there are no accounts or updater telemetry. The backend and MCP endpoint are unauthenticated on loopback; LAN sharing uses a PIN sent over plain HTTP unless you add a TLS proxy. App licence AGPL-3.0; default OmniVoice weights CC-BY-NC.
- What are the best VoiceStudio alternatives?
- The closest alternatives we have reviewed are audio.cpp, VideoAgent and ComfyUI.
Answers checked Oct 28, 2026.