VoiceStudio：本地优先的 AI 声音工作室，覆盖克隆、配音与转写 VoiceStudio: A Local-First AI Voice Studio for Cloning, Dubbing, and Transcription
开源项目 VoiceStudio 将声音克隆、语音合成、视频配音、有声书制作和语音转写整合到桌面端，并默认把数据与生成文件保存在本机。 Open-source VoiceStudio combines voice cloning, speech synthesis, video dubbing, audiobook production, and transcription in a desktop application that keeps data and generated files on the local machine by default.
VoiceStudio 是一款以本地运行为优先的开源 AI 语音制作平台。用户可以使用一段简短的参考音频克隆说话人的声音，也可以根据年龄、口音、音高、风格和表达方式等描述设计新声音。
面向多种语音工作流
其视频配音流程包括转写、翻译、保留说话人、合成和导出，适合需要在多个语言版本之间处理视频的用户。项目还支持多角色故事、有声书制作、EPUB/PDF 导入、章节渲染、.m4b 导出、系统级听写、声音与背景分离、说话人识别以及批量任务处理。
引擎选择决定实际能力
项目内置或接入 16 个 TTS 引擎和 11 个 ASR 引擎。可用语言、声音克隆能力、平台兼容性和资源需求取决于所选模型。用户可以通过模型目录安装、删除、选择和切换引擎，因此在部署前需要根据目标语言、硬件和模型授权情况进行核对。
本地隐私与合成音频标识
默认情况下，声音、项目、设置、转写结果和生成文件都会留在本机。远程工作节点、外部 ASR 服务和分析功能需要主动开启。VoiceStudio 还集成 AudioSeal 水印，用于检测合成音频，并明确建议用户只有在获得说话人授权后才能进行声音克隆。
桌面端与服务端架构
桌面端采用 Tauri v2 与 Rust，界面使用 React 和 Vite，并通过 IPC、本机 HTTP、SSE 和 WebSocket 与后端通信。运行在 3900 端口的 FastAPI 后端负责 REST 路由、流式服务、TTS/ASR 引擎注册、配音和长音频处理，同时提供兼容 OpenAI 的音频 API 与 MCP 服务。SQLite 和 Alembic 将应用状态保存到 omnivoice_data/。
部署入口
项目为 macOS 13.3+ Apple Silicon、Windows 10/11 x64 和 Linux x86_64（glibc 2.39+）提供预构建安装包。Docker 配置覆盖 CUDA、ROCm、CPU 以及仅运行 GPU Worker 的场景。首次启动会创建受管理的 Python 环境并下载默认模型，后续启动会复用已有环境和模型。源码运行需要 Node 20+ 或 Bun，以及 Python 3.11+。
声音克隆应建立在说话人明确授权的基础上；本地运行可以减少数据外传，但不会替代合规与授权判断。
VoiceStudio is an open-source AI voice production platform designed with local execution as a priority. Users can clone a speaker's voice from a short reference recording or design a new voice using descriptions such as age, accent, pitch, style, and delivery.
Workflows Beyond Text-to-Speech
Its video dubbing workflow covers transcription, translation, speaker preservation, synthesis, and export. The project also supports multi-character stories, audiobook production, EPUB/PDF import, chapter rendering, .m4b export, system-wide dictation, voice and background separation, speaker identification, and batch processing.
The Selected Engine Determines the Details
The project includes or connects to 16 TTS engines and 11 ASR engines. Available languages, voice-cloning capabilities, platform support, and resource requirements depend on the selected models. Users can install, remove, select, and switch engines through the model catalog, so deployment should be checked against target languages, hardware, and model licensing.
Local Privacy and Synthetic-Audio Detection
By default, voices, projects, settings, transcription results, and generated files remain on the local machine. Remote workers, external ASR services, and analytics must be enabled explicitly. VoiceStudio also integrates AudioSeal watermarking to help detect synthetic audio and advises users to obtain speaker authorization before cloning a voice.
Desktop and Backend Architecture
The desktop client uses Tauri v2 and Rust, with a React and Vite interface. It communicates with the backend through IPC and local HTTP, SSE, and WebSocket channels. The FastAPI backend, running on port 3900, handles REST routes, streaming services, TTS/ASR engine registration, dubbing, and long-audio processing. It also provides an OpenAI-compatible audio API and MCP service. SQLite and Alembic store application state in omnivoice_data/.
Installation Options
Prebuilt installers are provided for macOS 13.3+ on Apple Silicon, Windows 10/11 x64, and Linux x86_64 with glibc 2.39+. Docker configurations cover CUDA, ROCm, CPU, and GPU-worker-only deployments. On first launch, VoiceStudio creates a managed Python environment and downloads the default models; later launches reuse the existing environment and models. Running from source requires Node 20+ or Bun and Python 3.11+.
Voice cloning should be based on explicit speaker authorization. Local execution can reduce data transfer, but it does not replace compliance and consent decisions.
来源
- GitHub Trending · 09-04 08:57