timoncool/speech-studio
OFFLINE AI VOICE STUDIO IN RUST
Local app for speech synthesis and recognition with voice cloning and speaker diarization. Works offline, downloads models automatically, supports MCP for AI agents.
// KEY FEATURES
- TTS with built-in and cloned voices, EPUB/FB2/PDF support
- ASR with diarization: who said what and when, word-level cutting
- MCP server and OpenAI-compatible API for agent control
- Native Rust+C++, CUDA/Vulkan/Metal, models download in one click
// DEEP DIVE: HOW IT WORKS
The app is written in Rust with native C++ engines (GGUF/ONNX) for CUDA, Vulkan and Metal. No Python, Torch or CUDA wheels — the whole stack compiles into a single executable. On first run it downloads models, engines, CUDA/VC++ runtimes and ffmpeg itself; if a model already exists in another of the author's projects, it takes it from there. Works in three modes: Speech — voicing text, dialogs, podcasts and books; Listening — recognition with diarization, voicing verification against text, voice creation; Live — voice conversation with an AI agent via MCP.
Stands out for full locality: texts, recordings and voiceprints never leave the computer. Cloud voices (OpenRouter, Gemini, fal.ai, Runware) are optional, off by default, require a key and confirmation before first call. Has a built-in MCP server on port 8797 — agents like Claude Code or Cursor control the studio through the same functions as the UI. Also provides an OpenAI-compatible API (/v1/audio/speech, /v1/audio/transcriptions, /v1/models) for tools expecting OpenAI.
Primary target platform is NVIDIA GPU (CUDA); other backends (Vulkan, Metal, CPU) are less tested. Requires Windows, macOS (Apple Silicon) or Linux. MCP and API need a running app instance; microphone in Live mode is enabled only by the user. Cloud features require separate API keys. Distributed as a ready binary from GitHub releases.
// USE CASES
Best for: For audiobook, podcast, voice-over creators and AI agent developers
- Create audiobook from EPUB with chapters, characters and cover in M4B
- Transcribe meeting recording with speaker labels and timestamps
- Voice video script into subtitle cues with background music
// QUICK START
claude mcp add --transport http speech-studio http://127.0.0.1:8797/mcpCommands are taken from the repository README — check the current version on GitHub before running.
New repositories every 30 minutes
REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.
Join on Telegram