VibeVoice-RS

Rust implementation of VibeVoice text-to-speech with voice cloning and multi-speaker synthesis.

Features

• High-quality text-to-speech synthesis
• Voice cloning from audio samples
• Multi-speaker dialogue synthesis
• GPU acceleration (Metal/CUDA)
• Streaming audio generation (realtime model)

Crate Structure

Requirements

• Rust 1.85+
• HuggingFace account and token
• GPU recommended (Metal on Apple Silicon, CUDA on NVIDIA)

Setup

HuggingFace Token (Required)

# Create the cache directory
mkdir -p ~/.cache/huggingface

# Paste your token (get from https://huggingface.co/settings/tokens)
echo "hf_yourTokenHere" > ~/.cache/huggingface/token

# Secure it
chmod 600 ~/.cache/huggingface/token

Known Issues

Apple / Metal

Very long inputs may run over the buffer.

Error: Metal error Failed to create metal resource: Buffer

PyTorch MPS uses optimized SDPA (Scaled Dot Product Attention) that doesn't materialize the full attention matrix. Candle has flash attention but it's CUDA-only, not available for Metal.