# Overview

Introduction to VoiceStudio, its core capabilities, and how local processing works.

VoiceStudio is an open-source, local-first speech platform for voice cloning, voice design, video dubbing, dictation, transcription, and audiobook creation in 646 languages. It runs entirely on your own hardware, requiring no cloud accounts, external APIs, or internet connection for daily operations. It serves as a fully local alternative to commercial voice APIs, giving you complete control over your voice data and workflows.

## Core Capabilities

VoiceStudio provides workspaces tailored for multiple audio and speech tasks:

- **Voice Cloning & Gallery:** Create zero-shot voice clones using a clean reference recording as short as 3 seconds. Once created, manage, categorize, and preview your custom voices inside the Voice Gallery. Learn more in [Voice Cloning & Gallery](/guides/voice-cloning).
- **Voice Design:** Design entirely new custom voices using descriptive text and tags (such as `"female, whisper"`). Learn more in [Voice Design](/guides/voice-design).
- **Video Dubbing:** Automatically translate, transcribe, dub, and time-align video files while maintaining voice characteristics across languages. Learn more in [Video Dubbing](/guides/video-dubbing).
- **Audiobooks & Stories:** Generate longform multi-chapter content, featuring inline voice assignment, cue-sheet exports, and generation caching. Learn more in [Audiobooks & Stories](/guides/audiobooks).
- **Dictation & Transcription:** Dictate text on the fly using a global floating widget, or transcribe audio files with high-precision models like WhisperX. Learn more in [Dictation & Transcription](/guides/dictation).
- **Workflows & Batch Jobs:** Process massive workloads with automated generation queues, watch folders, and reusable generation recipes. Learn more in [Workflows & Batch Jobs](/guides/workflows).
- **Local API & MCP:** Integrate VoiceStudio directly into other local applications, custom scripts, or AI agent frameworks using an OpenAI-compatible web API and the Model Context Protocol (MCP). Learn more in [AI Agents & MCP](/guides/mcp) and the [Local API Reference](/reference/api).

> **Note:** By default, speech generation is powered by **OmniVoice** (k2-fsa/OmniVoice), but VoiceStudio is extensible and supports a variety of local speech engines such as IndexTTS, MOSS-TTS, dots.tts, and Confucius4-TTS.

## How Local Processing Works

VoiceStudio operates with a local-first architecture. When you click generate, transcribe, or dub, all processing is executed directly by the VoiceStudio backend running on your physical machine. 

### Hardware Acceleration

VoiceStudio automatically detects your hardware configurations to optimize performance:

- **NVIDIA GPUs (Windows & Linux):** Leverages CUDA acceleration, half-precision (`fp16`) inference, and `torch.compile` optimization. High-VRAM systems (16 GB+) can parallelize multiple concurrent generations.
- **Apple Silicon (macOS):** Utilizes Metal Performance Shaders (MPS) and MLX frameworks for fast, energy-efficient generation on Apple's unified memory.
- **CPU Fallback:** On older PCs, Intel Macs, or machines without a dedicated GPU, VoiceStudio automatically falls back to CPU execution. Synthesis remains fully functional but runs at a slower rate.
- **Windows on ARM:** Runs experimentally with a native ARM64 UI and an emulated x64 Python backend under CPU-only mode.

### Active Memory & VRAM Management

Speech models are large and can heavily consume GPU and system memory. VoiceStudio features intelligent memory systems to keep your hardware responsive:

- **Lazy Loading:** Models are loaded into memory only when they are first called, which typically adds an 8-second delay to the initial generation.
- **Active Offloading:** On machines with limited memory, VoiceStudio can automatically offload the active text-to-speech (TTS) model to system RAM when a transcription task requires video processing, or after a period of inactivity.
- **Manual Flushing:** If you hit memory ceilings or require memory for other local applications, you can use the **Flush** utility in the top toolbar to unload inactive models and clear cache fragments.

For more details on setting up your system and configuring hardware parameters, refer to [Installation & Setup](/getting-started) and [Models & Hardware](/guides/models).