Does VoiceStudio require payment or login?

The app is open source (AGPL-3.0) and does not force an account for local use. First run downloads models to disk; remote services and analytics need explicit consent.

Is audio uploaded by default?

No. Cloning, synthesis, and export stay local unless you enable remote or analytics features described in the app.

Does VoiceStudio run fully on Intel Macs?

No. Official macOS docs list Intel Macs as UI-only without the local inference backend. Use Apple Silicon or Windows/Linux for full features.

How long should clone reference audio be?

OmniVoice docs recommend about 3–10 seconds of clean single-speaker audio. A matching transcript helps; otherwise an installed ASR model can transcribe the clip.

How is this different from ElevenLabs?

VoiceStudio runs models on your hardware—better for privacy, offline batches, and engine choice. Hosted services trade hardware ops for per-minute billing and vendor custody of audio.

How do I export a video dub?

Use the Dubbing workspace: import media, assign voice profiles, render lines, then export the mix or replacement track per the Dubbing docs.

What can VoiceStudio do? Local AI voiceover, voice cloning, and video dubbing in a few steps

Voiceovers for tutorials, shorts, or product demos often take longer than the edit: talent, revisions, and timeline retakes add up. VoiceStudio bundles cloning, synthesis, transcription, and timed video dubbing in one Electron desktop app. Inference runs locally by default—no mandatory account and no cloud upload unless you opt in. The project is AGPL-3.0; docs and downloads live at voicestudio.sh. Features below follow the main branch documentation.

Unlike hosted APIs such as ElevenLabs, compute stays on your hardware. Reference clips, WAV exports, and batch jobs sit on disk. If you already validated other speech stacks on a Mac, see Qwen3-TTS Mac checks for a similar “download weights, then listen” workflow.

646

Language coverage in docs (check Model Catalogue)

3–10 s

Typical OmniVoice clone reference length

24 kHz

Default OmniVoice mono output rate

What VoiceStudio does

The README groups capabilities into Create (clone or design voices), Produce (timed dubs, audiobooks, batches), and Connect (local API, MCP, watch folders, recipes). Typical jobs:

Voice cloning: a few seconds of clean speech become a reusable profile.
Voice design: synthesize new timbres without a reference clip (UI varies by release).
Video dubbing: import video or audio, generate speech aligned to cues, export mixes.
Dictation & transcription: floating dictation plus optional whisper.cpp ASR in Model Catalogue.
Audiobooks / batch: long scripts split into chapters or files.

Model Manager lists alternate engines. OmniVoice ships as the default; low-VRAM or CPU-only setups can switch to lighter engines (KittenTTS, Supertonic, etc.). Profiles do not automatically transfer across engines.

License and commercial use

AGPL-3.0 has copyleft implications when you offer modified versions as a network service. Local personal use is usually straightforward; embedding VoiceStudio in a SaaS product needs legal review—not legal advice here.

OmniVoice defaults and hardware

OmniVoice (k2-fsa/OmniVoice) powers default TTS and cloning: zero-shot, broad language coverage, 24 kHz mono output. Weights load lazily on first use and are shared with dubbing and dictation.

GPU: docs recommend about 6 GB VRAM on a dedicated GPU; 4 GB cards may warn or run very slowly. Apple Silicon uses MPS with unified memory.
CPU: supported but slow for long clones or full dub passes; Windows installs a smaller CPU PyTorch build when no NVIDIA driver is present.
Reference audio: 3–10 seconds, single speaker, minimal noise; a matching transcript stabilizes conditioning. Missing text can be filled by an installed ASR model.

Creator recording reference audio at a microphone for VoiceStudio cloning — Clone quality starts with the reference recording, not slider tweaks.

Install on macOS

The macOS app is Electron-based; legacy Tauri builds are deprecated. Follow docs/install/macos.md for packages and channels.

Download the Apple Silicon build from GitHub Releases or the site; drag to Applications.
Allow the developer in System Settings if Gatekeeper blocks the first launch.
First synthesis triggers model downloads; keep disk and network headroom.
Intel Macs are UI-only—local inference is unsupported. Use Apple Silicon or another OS for full backend features.

First clone in a few steps

Open the Voice Cloning workspace.
Create a profile; upload 3–10 s of reference audio; paste a transcript if you have one.
Enter target text, keep OmniVoice selected, generate.
Preview the 24 kHz clip; save the profile for dubbing and batch jobs.

Clone only voices you are allowed to use; impersonation may violate law and platform rules.

Video dubbing workspace

Dubbing solves timing: import media, split lines, assign profiles per character, render aligned speech. Official docs cover multi-speaker exports under docs/. Tips:

Lock picture edits before long TTS passes.
One profile per role; split narration from dialogue tracks.
Upsample in a DAW if your NLE expects 48 kHz.

Dictation, API, agents

Dictation floating widget for drafts.
Transcription via whisper.cpp or other ASR models.
Local API / MCP for agents; optional skill pack via npx skills add debpalash/VoiceStudio.
Recipes & watch folders for repeatable batch pipelines.

Apple Silicon and Cloud Mac

Long dubbing batches want sustained GPU time. A 16 GB+ Apple Silicon Mac—or a Cloud Mac you reach over SSH/VNC—keeps your laptop free for editing. This pairs with a local AI gateway: agents draft scripts; VoiceStudio renders audio locally.

FAQ

Short answers: open source on your hardware; offline by default; Intel Mac UI-only; 3–10 s references; dub in the Dubbing workspace. Structured FAQ is in the page head.