Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# MeanFlowSE — One-Step Generative Speech Enhancement
|
| 2 |
|
| 3 |
**MeanFlowSE** is a conditional generative approach to speech enhancement that learns average velocities over short time spans and performs enhancement in a single step. Instead of rolling out a long ODE trajectory, it applies one backward-in-time displacement directly in the complex STFT domain, delivering competitive quality at a fraction of the compute and latency. The model is trained end-to-end with a local JVP-based objective and remains consistent with conditional flow matching on the diagonal—no teacher models, schedulers, or distillation required. In practice, 1-NFE inference makes real-time deployment on standard hardware straightforward.
|
|
@@ -216,4 +221,4 @@ We gratefully acknowledge **Prof. Xie Chen’s group (X-LANCE Lab, SJTU)** for t
|
|
| 216 |
|
| 217 |
|
| 218 |
**Questions or issues?** Please open a GitHub issue or pull request.
|
| 219 |
-
We welcome contributions — from bug fixes to new backbones and front-ends.
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
datasets:
|
| 4 |
+
- JacobLinCool/VoiceBank-DEMAND-16k
|
| 5 |
+
---
|
| 6 |
# MeanFlowSE — One-Step Generative Speech Enhancement
|
| 7 |
|
| 8 |
**MeanFlowSE** is a conditional generative approach to speech enhancement that learns average velocities over short time spans and performs enhancement in a single step. Instead of rolling out a long ODE trajectory, it applies one backward-in-time displacement directly in the complex STFT domain, delivering competitive quality at a fraction of the compute and latency. The model is trained end-to-end with a local JVP-based objective and remains consistent with conditional flow matching on the diagonal—no teacher models, schedulers, or distillation required. In practice, 1-NFE inference makes real-time deployment on standard hardware straightforward.
|
|
|
|
| 221 |
|
| 222 |
|
| 223 |
**Questions or issues?** Please open a GitHub issue or pull request.
|
| 224 |
+
We welcome contributions — from bug fixes to new backbones and front-ends.
|