DARe-VQ

This repository provides the official pretrained checkpoints of DARe-VQ.

DARe-VQ is a large-codebook single-vector-quantization framework designed for high-quality speech representation learning. It improves the training stability and codebook utilization of large EMA-based VQ models, enabling scalable discrete speech tokenization.

For the method details, implementation, and experimental results, please refer to the official GitHub repository:

https://github.com/XYGuo1996/DARe-VQ

Model Details

• Model type: Neural speech tokenizer / vector quantizer
• Domain: Speech representation learning
• Sampling rate: 24 kHz
• Training data: LibriTTS
• Codebook sizes: 16K and 131K
• Token rate: 75 Hz

Available checkpoints:

Model	Description
DARe-VQ-16K	16,384-entry single VQ codebook
DARe-VQ-131K	131,072-entry single VQ codebook

Intended Uses

DARe-VQ can be used for:

• Discrete speech token extraction
• Speech representation learning
• Neural speech codec research

Usage

Please refer to the GitHub repository for installation instructions and inference examples:

https://github.com/XYGuo1996/DARe-VQ