GA-AF-FlowSE
Forward-Process Reinforcement Learning for Flow-Matching Speech Enhancement
Overview
This repository hosts the released checkpoints for GA-AF-FlowSE, including our GA-AF, AdvantageFlow (AF), and controlled GRPO models.
All models are built on top of FlowSE.
For method details, training configurations, evaluation, and reproduction instructions, please see the GitHub repository.
Checkpoints
| Model | Checkpoint |
|---|---|
| Base | base_sft20k/flowse_sft20k.pt |
| AF | af/af_libritts_step005000.pt |
| GA-AF | gaaf/gaaf_libritts_step005000.pt |
| GRPO | grpo/grpo_libritts_step005000.pt |
Download
hf download spzhang7/GA-AF-FlowSE gaaf/gaaf_libritts_step005000.pt
Or with Python:
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download(
repo_id="spzhang7/GA-AF-FlowSE",
filename="gaaf/gaaf_libritts_step005000.pt",
)
Upstream FlowSE
The original FlowSE checkpoint is not redistributed in this repository.
Please download it from the official repository:
https://huggingface.co/flowse/wenetspeech4tts_Premium.pt.tar
Code
Training, inference, evaluation, and full reproduction instructions are available at:
https://github.com/spzhang7/GA-AF-FlowSE
Citation
@misc{zhang2026forwardprocess,
title = {Forward-Process Reinforcement Learning for Flow-Matching Speech Enhancement},
author = {Shuaipeng Zhang and Hang Chen and Jun Du and Qing Wang and
Sabato Marco Siniscalchi and Zihao Quan and Zijing Cai and Boyuan Dong},
year = {2026},
note = {GA-AF-FlowSE},
url = {https://github.com/spzhang7/GA-AF-FlowSE}
}
Please also cite FlowSE when using the upstream model or code.