modelscope/FunCodec

Implements multiple neural codec architectures (EnCodec, FreqCodec) with configurable quantization schemes and bitrates (250-16000 bps), integrated with ModelScope and Hugging Face for model distribution. Built on Kaldi-style data organization (`wav.scp`) supporting both waveform and ark file formats, enabling training on custom datasets alongside downstream tasks like codec-based TTS (LauraTTS) that outperforms VALL-E. Provides end-to-end batch inference pipelines with distributed training via torchrun and recipe-based workflows for reproducibility across multiple open-source corpora.

442 stars. No commits in the last 6 months.