OmniVoice Studio is an open-source desktop application that positions itself as the local ElevenLabs alternative: zero-shot voice cloning, real-time dictation, audiobook production, and cinematic video dubbing, all running on your own hardware with no accounts, no API keys, and no cloud. Built by developer Palash Deb (debpalash) and released under AGPL-3.0, the project has climbed past 8,500 GitHub stars and 1,300 forks since its April 2026 debut, riding a wave of interest in privacy-first voice AI. Its headline number is language coverage: 646 languages, driven by the k2-fsa OmniVoice model it ships as its default engine. ## A Studio, Not a Model The project's key design decision is that it is an application layer, not another TTS model. Under the hood it orchestrates 14 TTS engines — OmniVoice (default, 600+ languages), CosyVoice 3, GPT-SoVITS, VoxCPM2, MOSS-TTS-Nano, KittenTTS, MLX-Audio, Sherpa-ONNX, and lazy-installed heavyweights like IndexTTS 2, Supertonic 3, MOSS-TTS-v1.5 (8B), dots.tts, and Confucius4-TTS — plus 10 ASR engines for transcription. A single settings picker switches the active engine everywhere synthesis happens, and adding a new engine means subclassing a `TTSBackend` in roughly 50 lines. This aggregator approach means the app improves automatically as the open TTS ecosystem does. ## Cloning, Dubbing, and Audiobooks Voice cloning is zero-shot from a 3-second clip, with voice design controls (gender, age, accent, pitch, style, dialect) that go beyond what ElevenLabs exposes. The video dubbing pipeline is fully local and surprisingly complete: Demucs-powered vocal isolation splits speech from the music bed, Pyannote + WhisperX diarization identifies who said what, and cloned voices re-speak the dialogue in the target language while the background audio survives intact. A batch queue accepts dozens of videos with per-job progress. The audiobook editor imports EPUB and PDF, supports multi-voice stories, and exports .m4b — a feature ElevenLabs simply doesn't offer. Cloned voices export as portable .ovsvoice bundles carrying identity plus an AudioSeal (Meta) invisible watermark that survives compression. ## Engineering for Real Machines The project is notably serious about the unglamorous parts of local AI. GPU support auto-detects CUDA, Apple Silicon MPS, and ROCm (Linux, opt-in), with automatic TTS-to-CPU offload on GPUs with 8 GB VRAM or less. Engine routing runs a preflight GPU check per engine and refuses silent CPU fallbacks. A built-in self-check suite, error journal, and scrubbed diagnostic bundles make bug reports actionable. There is an OpenAI-compatible API, a WebSocket streaming path with no text-length cap, a remote-backend mode (Tailscale-friendly, bearer auth) so a laptop UI can drive a GPU box, and an MCP server so Claude, Cursor, or any MCP client can drive synthesis directly. ## The Privacy and Cost Pitch The comparison the README leads with is blunt: ElevenLabs charges $5–$330 per month with per-character billing and processes audio server-side; OmniVoice Studio is free, unlimited, and nothing leaves the machine. For dictation — where the microphone content is by definition the most sensitive data stream on a computer — the local-LLM transcript cleanup and optional echo cancellation make it a genuine daily-driver replacement for cloud dictation tools. ## Trade-offs and Limitations The project is explicitly in active beta and warns that releases can break; running from source gets the newest fixes. The AGPL-3.0 license requires a commercial license for proprietary embedding, which matters for companies. Hardware constraints are real: macOS Intel cannot run the local backend at all (PyTorch dropped Intel-Mac wheels), Windows AMD GPUs fall back to CPU-only, and the heavyweight engines want 9–16 GB of downloads in isolated venvs. Quality ceilings are inherited from whichever open engine is active, and cloning-dependent features fail up front (by design) when a clone-less engine is selected. As with all voice cloning tools, the deepfake risk sits with the user — the AudioSeal watermarking is a mitigation, not a guarantee. ## Who Should Use This OmniVoice Studio fits creators producing audiobooks and dubbed video without per-character bills, privacy-conscious users who want dictation and voice AI that never phones home, and developers who want one API over the entire open TTS ecosystem. Anyone who migrated away from the now-archived Real-Time-Voice-Cloning project even gets a dedicated migration guide. It is the most complete open-source answer yet to the question of what ElevenLabs would look like as a local app.