VoiceStudio

VoiceStudio ships native desktop apps for macOS (Apple Silicon and Intel), Windows x64, and Linux x64 (AppImage and deb packages), installed with one command via the installer script at voicestudio.sh/install. There is no mobile app. For headless deployment there is an official Docker image, ghcr.io/debpalash/voicestudio, published for linux/amd64 only with ROCm variants.

Yes. All speech models run locally and remote services are optional. Model downloads need one-time internet access, but cloning, dubbing, dictation, and transcription afterwards run on your own hardware. Usage analytics only sends data if you explicitly opt in, and the local API binds only to loopback by default.

Yes. It exposes a local HTTP and WebSocket API on port 3900 and an MCP server, so tools like Claude Code, Codex, and other MCP-compatible agents can start dictation, run text-to-speech, or request transcriptions. A well-known endpoint at 127.0.0.1:3902 describes all available capabilities, and the executable accepts dictation flags from the command line.

The default speech engine is k2-fsa OmniVoice; you can switch to CosyVoice 3, IndexTTS 2.5, MLX-Audio, GPT-SoVITS, KittenTTS, sherpa-onnx, and others. Transcription defaults to WhisperX with Faster-Whisper, MLX Whisper, PyTorch Whisper, NVIDIA Parakeet, Moonshine, FunASR, and any OpenAI-compatible server as alternatives. Engines are selected in-app per task.

Requirements vary by engine, but the project benchmarks on 16 GB Apple Silicon (M2) as a baseline. NVIDIA CUDA GPUs, Apple MLX/Metal, and AMD ROCm on Linux are auto-detected. The Docker headless image is CPU-only on non-GPU machines and requires linux/amd64; Apple Silicon Macs should use the native app rather than the container for GPU acceleration.