OmniVoice Studio

The pitch is on the homepage in a single line: "Same jobs, different terms." OmniVoice Studio lines its feature table up against ElevenLabs and argues the difference is where the audio lives, not what the tool does. On the vendor's own comparison, cloning is a 3-second clip either way; voice design goes further locally (gender, age, accent, pitch, style, dialect versus gender and age); audiobooks get a full editor rather than a blank; and dubbing moves from cloud-only to fully local. Those are the sorts of claims you can test in an afternoon, which is the point — this is a desktop app you install rather than a service you buy. The engineering choice worth calling out is engine coverage. Rather than one monolithic model, the studio ships 16 TTS and 11 ASR engines behind a single picker reachable with Cmd/Ctrl+E. VoiceStudio handles 600+ languages with clone-and-instruct, CosyVoice 3, VoxCPM2 and IndexTTS 2.5 cover the rest, and MLX-Audio targets Apple Silicon specifically. On the recognition side, WhisperX gives word-level timing across roughly 100 languages, Parakeet TDT runs at about 10x realtime on CPU, and FunASR adds inline diarization. That breadth means you are not locked to one model's accent or failure mode. It also means model downloads and disk usage are yours to manage — 10 GB minimum, 20 GB+ SSD recommended. Hardware is the honest dividing line. The system floor is 8 GB RAM, 4 GB VRAM and 10 GB disk, and the vendor is explicit that a GPU is optional: TTS auto-offloads to CPU and the pipeline runs, just slower. Recommended is 16 GB+ RAM and 8 GB+ VRAM (RTX 3060-class). On a laptop with integrated graphics you will get there, but you will feel it. If your work is realtime or high-volume, the GPU question decides whether this is pleasant or merely possible. The integration story is deliberately unglamorous and unusually useful. An OpenAI- and ElevenLabs-compatible API on localhost means existing SDK code can point at your machine with no rewrite; voice selected by your own cloned-profile IDs; model pins an engine per request. The MCP server exposes the studio to Claude, Cursor and any MCP client, so an agent can drive synthesis rather than you clicking through a UI. Docker deployment covers server use. Where it does not fit: teams that need shared projects, roles or collaborative editing — this is a single-user desktop studio. Anyone who wants a browser or mobile interface. Anyone who wants a vendor to handle setup and maintenance. And buyers chasing the last few percent of naturalness, where dedicated cloud vendors still have an edge. The governance angle is real. AGPL-3.0 means derivative works may need to be open-sourced — fine for individual creators and for engineers who intend to fork, a conversation for commercial embedding. In a 2026-08-28 blog post, Debnath argues that "local-first" is too blunt a label and that useful trust boundaries are more specific: VoiceStudio, Opal and Bootable each keep different things under user