VoiceStudio Plus puts voice cloning, voice design and video dubbing in one desktop workspace. Five seconds of reference audio is enough to clone a voice — and everything runs on your own GPU. Only the first model download needs a network.
Cloning, design and dubbing each have their own screen with the parameters that matter for that job. Voices, models and takes all live in one project library.
Give it 5–15 seconds of clean speech and it clones that voice, ready to say anything you type.
No sample needed: describe gender, age, pitch, accent, dialect and style, and the engine builds a voice from scratch.
Import a video: it transcribes, splits by sentence and dubs each line into the target language, then exports a time-aligned result.
Most voice tools die at the install step. Here the runtime, the models and the mirrors are already wired up.
Cloning and synthesis happen on your machine. Reference audio, manuscripts and finished takes never reach a server — and it keeps working offline.
The default engine covers 646 languages; dubbing, transcription and translation across languages share one model foundation.
Chaptered rendering and an unattended batch queue, with each job keeping its own language, voice and background settings.
Point n8n, LiteLLM or your own scripts at the local service and drive speech, transcription and translation over the /v1 endpoints.
Multi-speaker recordings are split by speaker automatically, and a floating dictation widget turns speech into text in any app.
The runtime, every dependency and the required model come from our own mirrors — no direct huggingface.co or PyPI needed.
With an NVIDIA card it enables CUDA automatically, and more VRAM is faster; on our own machine an RTX 2080 Ti with 22 GB runs everything.
It also runs without an NVIDIA GPU — setup installs the CPU build of the inference library automatically, just much slower. Apple Silicon and AMD cards have no separate acceleration path yet.
About 221 MB, shell only. On first launch it sets up the Python runtime in the background: about 9.5 GB for CUDA, about 5 GB for CPU.
The required model adds about 3 GB. Leave 15 GB or more free after installation; the runtime folder can be moved to another drive from Settings.
It installs either the CPU or CUDA inference libraries depending on your machine, then fetches the required model, with progress shown in the app.
Every download resumes — close the app mid-way and it picks up next time. The required model is mirrored by us, so mainland China connects directly.
No. Cloning, design, dubbing and transcription all run locally; reference audio and output only touch your disk.
Only the first model download needs the network; afterwards it works offline. In-app usage analytics are off by default and need your explicit consent.
VoiceStudio is an open-source project by Palash Debnath and contributors (AGPL-3.0). This build is based on it, adapted for Chinese-language work and mainland networks.
The free part keeps every creative feature; additional production-workflow tools ship with the Plus edition. The in-app Plus page lists exactly what is included.
The default engine covers 646 languages including Chinese, English, Japanese, Korean and the major European languages. Voice design also offers Chinese dialects such as Sichuanese and several English accents.
Settings has the runtime log and diagnostics. Start with the error tail under Settings → Logs; model problems usually show up on the Models page as a missing file.
If it is still unresolved, send the log tail together with your system information.
Windows 10 / 11 · 64-bit · The installer ships the shell; the runtime is configured on first launch
Installer about 221 MB · runtime about 9.5 GB · required model about 3 GB
Double-click to install. Installing over an existing version keeps your data folder. SHA-256：d466bacc50e0c8aa993de638167e8560ec36b028601af799a2351713fd51c9a0