AI-generated imageThis article was generated and researched by Arthur, AiGENCY’s persistent-memory AI. It is fact-checked against the cited sources, but may still contain errors.
Is there really a free, open-source ElevenLabs replacement that runs on your own machine? For UK readers on 13 September 2026, the short answer is: VoiceStudio by Palash Debnath is the closest thing to it. Voice cloning from a three-second clip, video dubbing, dictation, transcription and audiobooks, all local, no per-character billing. But free software is not free computing, it is an active beta, and the biggest language-count claim needs a pinch of salt. Here is the honest skim.
What is VoiceStudio?
VoiceStudio, previously called OmniVoice-Studio, is an open-source desktop app by developer Palash Debnath that pitches itself directly as the open-source ElevenLabs alternative. The tagline on the repository says it plainly: real-time dictation, zero-shot voice cloning and cinematic video dubbing, all on your desktop, with no accounts, no API keys and no cloud. The code is published under the GNU Affero General Public License v3.0, and the project page lives at https://github.com/debpalash/VoiceStudio with downloads and docs at https://voicestudio.sh.
How popular is it? Very, and moving fast. The repository shows roughly ten thousand stars and over one and a half thousand forks, with social posts variously claiming nineteen to twenty-four thousand stars — star counts bounce with caching and the project rename, so read it as strong momentum rather than an exact figure. There have been dozens of releases, the latest around v0.5.2, and the maintainer has over a thousand commits. Solo-developer velocity at this scale is rare and worth respecting, but it also means support is a Discord server and a GitHub issues page, not a service desk.
What can it actually do?
What does the feature list cover? Seven workloads in one studio: voice cloning, voice design from scratch by gender, age, accent, pitch and emotion, a gallery of ready-made voices, video dubbing end to end with transcription, translation, re-voicing and timing, multi-voice story and audiobook creation with EPUB import, transcripts from audio or video, and real-time dictation with local language-model cleanup.
How does the engine system work? Instead of one built-in voice model, VoiceStudio is a shell around fourteen to sixteen text-to-speech engines and eleven speech-recognition engines, switchable in Settings. The default engine covers the headline language catalogue, with others such as CosyVoice 3, GPT-SoVITS, IndexTTS 2 and Supertonic opt-in or lazy-installed. A compatibility matrix runs a per-engine GPU preflight check, so it tells you what your machine can actually run rather than failing silently onto the CPU.
What about the 646 languages claim? This is the catalogue count across engines, not equal quality in every language. English, Bengali, Hindi and major European languages get the showcase demos, including a full Bengali video dub in the screenshots. Smaller languages may work technically while sounding clearly synthetic. Treat it as coverage breadth, which is genuinely useful for UK community organisations serving many mother tongues, not as 646 studio voices.
How do you install it, and what hardware do you need?
How do you get it running? The fastest route is the one-line installer for macOS, Linux and Windows Subsystem for Linux: curl -fsSL https://voicestudio.sh/install | sh. There are also signed installers for Apple Silicon Macs, Intel Macs, Windows and Linux, or you can run from source with git clone plus bun install and bun run desktop. Mac users should expect the one-time Gatekeeper approval dance of right-click then Open.
What machine do you actually need? This is where free meets physics. Eight gigabytes of RAM is the stated minimum and will block some engines. Apple Silicon Macs are the smoothest consumer path via MPS and MLX acceleration. NVIDIA GPUs with CUDA are the best performer on desktop. Intel Macs cannot run the local backend, AMD graphics on Windows falls back to CPU-only, and CPU-only synthesis of long audiobooks is slow enough to plan around. Model weights are gigabyte-scale downloads — one engine pull is around 2.4 gigabytes — plus per-engine virtual environments, and the catalogue now shows a disk breakdown before you install.
What trips people up in setup? The troubleshooting docs list the greatest hits: WhisperX dependency breaks, PyTorch version mismatches, gated Hugging Face weights such as speaker-diarization models needing a free token and a consent click, and antivirus software corrupting large model downloads. None of this is unusual for local AI, but it is a world away from signing up to a cloud dashboard. Budget an evening, not ten minutes.
How does it compare with ElevenLabs?
What does the project's own comparison table admit? VoiceStudio matches ElevenLabs on three-second zero-shot cloning and beats it on voice-design controls and built-in audiobook tooling, while ElevenLabs keeps the edge on polished cloud API, premade voice library and zero-setup onboarding. Pricing is the stark row: ElevenLabs subscriptions run to hundreds per month with per-character billing, while VoiceStudio the software is free and you supply the hardware and electricity.
Is it really as good? The repository FAQ says yes for cloning and dubbing, but that is the author's own answer, not an independent verdict. The closest thing to a review, a MarkTechPost walkthrough, largely restates the repository with architecture detail rather than running blind listening tests. No credible A/B shootout between the two has surfaced. For narration and dubbing drafts the gap looks small; for the most demanding commercial voice work the cloud incumbent still has the polish advantage.
When does local win outright? Three cases. Privacy: nothing leaves your machine, which matters for unreleased book manuscripts, client recordings and sensitive charity casework. Volume: a full audiobook that would cost real money per character in the cloud costs only your overnight render locally. Control: fourteen engines, watermarking options, an MCP server for agent clients and a local API mean tinkerers can wire it into anything.
What are the catches UK users should know?
Is it stable? It is labelled active beta and means it. The issues page shows real reports: generation-capacity busy errors, an AudioSeal watermark compile failure on macOS wasting half a minute per take, and individual engine failures. Run the latest release for steadier behaviour, run from source for the newest fixes, and do not schedule client work the same evening you install.
Can you use it commercially? The software licence is AGPL-3.0, which is free for personal and internal use but carries share-alike obligations if you offer it over a network, and the project notes a separate commercial licence for proprietary use. Voices, models and training data carry their own licences on top. If you are a UK business planning paid audiobook or dubbing output, read the licence files and the model cards before you invoice anyone.
What about cloning other people's voices? This is the serious paragraph. VoiceStudio embeds Meta's AudioSeal watermark, which survives compression and proves audio is AI-generated, but watermarking is provenance, not prevention. A three-second clip is enough to clone, and there is no evidence of a speaker-consent gate, liveness check or cloud verification step of the kind ElevenLabs operates. Local-first means nobody can stop misuse at the server, because there is no server. UK readers should know that non-consensual voice cloning can engage fraud, harassment and data-protection law. Clone your own voice, licensed voices and consenting collaborators, and keep records of consent for client work.
Who should try it this weekend?
Should you download it? If you are a writer wanting a private audiobook draft of your own book, a charity producing multilingual training audio, a developer wiring voice into a local agent stack, or a tinkerer with a decent GPU, yes — this is the best free route in its class right now. If you need guaranteed turnaround, a phone call to support, or commercial paperwork by Monday, pay for the cloud tool and revisit VoiceStudio at v1.0.
The bottom line: VoiceStudio does not quite delete ElevenLabs, but it deletes the bill for an enormous range of personal, creative and community work, at the price of setup effort and beta rough edges. For a UK audience watching every pound, that trade is increasingly the whole story of open-weight AI. Check the repository, check your hardware against the engine matrix, and go in with evening-project expectations.