OpenMOSS/MOSS-Audio-Tokenizer

Based on the README, here's a technical summary that goes deeper: Employs a 32-layer Residual Vector Quantizer with pure Causal Transformer blocks (1.6B parameters total) to compress 24kHz audio to 12.5Hz frame rate while supporting bitrates from 0.125–4kbps. Trained end-to-end without pretrained encoders or distillation, jointly optimizing encoder, quantizer, decoder, discriminator, and an LLM component for semantic alignment. Available in PyTorch on Hugging Face and ModelScope, with ONNX Runtime and TensorRT backends for deployment-ready inference without PyTorch dependencies.

162 stars.