Python-free VibeVoice Engine Released

vibevoice.cpp is a project that ports Microsoft's VibeVoice model to ggml/C++, enabling standalone execution during inference without Python, PyTorch, or vLLM.

Key Features:

• TTS (Text-to-Speech): 24kHz voice cloning possible with a 30-second reference clip (0.5B model).
• Long-form ASR (Automatic Speech Recognition): Supports a 7B parameter model with Diarization (speaker separation) functionality. Processes up to 17 minutes of audio at once and returns results in JSON format.

Supported Backends: Supports CPU, CUDA, Metal, Vulkan, and hipBLAS, enabling operation across various hardware environments.

Performance and Features:

• Recorded 0.41 RTF (Real-Time Factor) on a 68-second sample in a CUDA(Q4_K) environment.
• Unlike existing Python-based approaches, it has absolutely no Python dependencies, making it highly advantageous for embedding and local deployment.