VoiceStudio

"OmniVoice Studio" is the old name of VoiceStudio (debpalash/VoiceStudio; the old URL github.com/debpalash/OmniVoice-Studio redirects there). The rename page says it is the same project, repo, and API; "OmniVoice" now names the default engine and some old installer files. It is a desktop app with an Electron front end and a FastAPI backend, and everything runs on the machine. There is no account and no cloud, apart from updates and model downloads. Voice cloning is zero-shot, so no training runs per voice. The app ships 14 swappable TTS engines and defaults to OmniVoice. Checked 2026-10-17.

Not to be confused with the model itself, k2-fsa/OmniVoice, which VoiceStudio wraps.

Features

• Voice cloning from a short reference clip, voice design (describe gender, age, pitch, accent), saved Voice Profiles reusable across features.
• Video dubbing with timed speech, stories/audiobooks, batch jobs, watch folders, dictation widget, transcription.
• 14 swappable TTS engines in a Model Catalogue; OmniVoice is the default. A GGUF build of OmniVoice is offered for low-VRAM/CPU machines.
• Local OpenAI-compatible API on localhost:3900 (e.g. POST /v1/audio/speech) and an MCP server for agents.
• Non-verbal tags such as [laughter] and pronunciation correction (OmniVoice README).

Voice cloning approach and samples

Zero-shot: no training per voice. The OmniVoice engine doc says 3–10 seconds of clean speech is the sweet spot; a matching transcript improves conditioning (VoiceStudio can transcribe it with an installed STT model). With a supplied transcript the clip is capped at 20 s; without one, the app searches up to 75 s and picks the best 15 s window. Accepts WAV, MP3, M4A, FLAC, OGG, or a mic recording. Output is 24 kHz mono.

Vietnamese

Vietnamese (vi) is in the OmniVoice language list with about 8,482 hours of training data (VoiceStudio docs/languages.md). Cross-language cloning is advertised. Real Vietnamese quality (tones, regional accents, numbers) was not listened to for this note.

Hardware and platforms

• macOS (Apple Silicon, Metal), Windows, Linux (.dmg, .exe, .AppImage/.deb), Docker. Intel Mac: UI only, needs a remote backend.
• OmniVoice engine recommends a dedicated GPU with ≥6 GB VRAM; 4 GB cards page to RAM and become very slow. CPU works, slower (~5 GB disk for the CPU PyTorch build). Site says ~8 GB RAM is enough to start.
• Install: curl -fsSL https://voicestudio.sh/install | sh (macOS/Linux) or the PowerShell one-liner on Windows. First launch downloads multi-GB model weights.
• No account, no cloud; network only for updates and model downloads. Analytics only with consent.

Pricing and licensing

Important for commercial use:

• The app license does not cover model weights. VoiceStudio's LICENSE-NOTICE says the k2-fsa/OmniVoice model card identifies the pretrained weights as CC-BY-NC (non-commercial), and that audio_tokenizer/LICENSE carries Boson Higgs Audio 2 / Meta Llama community terms. On 2026-10-17 the HF model card had no license field, and audio_tokenizer/LICENSE was confirmed to be the Boson Higgs Audio 2 Community License. The CC-BY-NC statement was not re-confirmed on the current card; check before any commercial output.
• A paid Pro license does not grant model or voice rights. Clone only with the voice owner's permission.

Pros and cons

Pros : Fully local and private, free, cross-platform, very broad language list including Vietnamese, short 3–10 s samples, OpenAI-compatible local API and MCP, engine choice.

Cons : Needs a decent GPU for speed; multi-GB downloads. Default model's weight license is likely non-commercial, so commercial voiceover needs another engine or a license check. Zero-shot clones are generally less faithful than a fine-tuned PVC (not measured here). Young project (created April 2026), fast-moving.

See Voice Cloning Comparison.