tags:
- speech-enhancement
- speech-dereverberation
- diffusion
- consistency-distillation
- consistency-training
- real-time
- audio language:
- en pipeline_tag: audio-to-audio
ROSE-CD: Robust One-step Speech Enhancement via Consistency Distillation
Liang Xu, Longfei Felix Yan, W. Bastiaan Kleijn
IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2025 — Oral Presentation
Overview
ROSE-CD proposes a one-step consistency training (CT) framework for real-time speech enhancement. It accelerates inference by 54× compared to the 30-step teacher diffusion model while achieving superior performance across all objective metrics.
Available Checkpoints
| Directory | Description | Benchmark |
|---|---|---|
logs/CT_pesq5e-4_L2/ |
One-step CT model checkpoint | VoiceBank-DEMAND |
logs/teacher/ |
30-step Teacher model checkpoint | VoiceBank-DEMAND |
logs/CT_SE_EARS_pesq5e-4_L2/ |
One-step CT model checkpoint | EARS-WHAM (SE) |
logs/teacher_SE_EARS/ |
Teacher model checkpoint | EARS-WHAM (SE) |
logs/CT_Reverb_EARS_pesq5e-4_L2/ |
One-step CT model checkpoint | EARS-REVERB (Dereverb) |
logs/teacher_Reverb_EARS/ |
Teacher model checkpoint | EARS-REVERB (Dereverb) |
Enhanced audio outputs are provided under out/.
Performance
VoiceBank-DEMAND
| Model | Steps | PESQ (↑) | ESTOI (↑) | SI-SDR (↑) | SI-SIR (↑) | SI-SAR (↑) |
|---|---|---|---|---|---|---|
| Teacher | 30 | 2.89 | 0.86 | 16.7 | 26.7 | 17.6 |
| CT (Ours) | 1 | 3.47 | 0.87 | 19.2 | 29.2 | 20.0 |
EARS-WHAM (Speech Enhancement)
| Model | Steps | WER (↓) | PESQ (↑) | SI-SDR (↑) | ESTOI (↑) |
|---|---|---|---|---|---|
| SGMSE+ | 60 | 18.65% | 2.20 | 14.2 | 0.84 |
| ROSE-CD (Ours) | 1 | 18.19% | 2.81 | 15.3 | 0.85 |
EARS-REVERB (Dereverberation)
| Model | Steps | WER (↓) | PESQ (↑) | ESTOI (↑) |
|---|---|---|---|---|
| SGMSE+ | 60 | 17.32% | 1.95 | 0.76 |
| ROSE-CD (Ours) | 1 | 15.78% | 2.69 | 0.82 |
Download
pip install huggingface_hub
huggingface-cli download LIANGXU123/ROSE-CD --local-dir ./
# To download a specific checkpoint: --include "logs/CT_pesq5e-4_L2/*"
Place the downloaded logs/ directory inside the code repository and follow the evaluation instructions therein.
Citation
@inproceedings{xu2025robust,
title={Robust One-step Speech Enhancement via Consistency Distillation},
author={Xu, Liang and Yan, Longfei Felix and Kleijn, W Bastiaan},
booktitle={2025 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA)},
pages={1--5},
year={2025},
organization={IEEE},
doi={10.1109/WASPAA66052.2025.11230988}
}
License
This project is released under the MIT License.