spellbrush/AliasingFreeNeuralAudioSynthesis

Aliasing-Free Neural Audio Synthesis

This is the official Hugging Face model repository for the paper "Aliasing-Free Neural Audio Synthesis", which is the first work to achieve simple and efficient aliasing-free upsampling-based neural audio generation in the entire field of neural vocoders and codecs.

For more details, please visit our GitHub Repository.

Model Checkpoints

This repository contains the following checkpoints:

Model Name	Directory	Description
Pupu-Vocoder_Small	./pupuvocoder/*	14M parameter small version of Pupu-Vocoder.
Pupu-Vocoder_Large	./pupuvocoder_large/*	122M parameter large version of Pupu-Vocoder.
Pupu-Codec_Small	./pupucodec/*	32M parameter small version of Pupu-Codec.
Pupu-Codec_Large	./pupucodec_large/*	119M parameter large version of Pupu-Codec.

How to use

You need to put the pretrained models in:

  AliasingFreeNeuralAudioSynthesis/experiments

of our official repository, and then follow the instructions written in the repository to resume, finetune, and inference our pretrained checkpoints.

Citation

@article{afgen,
  title        = {Aliasing Free Neural Audio Synthesis},
  author       = {Yicheng Gu and Junan Zhang and Chaoren Wang and Jerry Li and Zhizheng Wu and Lauri Juvela},
  year         = {2025},
  journal      = {arXiv:2512.20211},
}