Dereverberation

[ARKEYTECT]

Great work on the model. One quick question, it seems the published model are not trained with dereverberation. Is it planned in the future to release models with reverb cleaning capabilities?

Thanks

[aask1357]

Hello. Thank you for your interest! I think adding dereverberation is a good idea. I expect to release it around May. BTW, would 16kHz versions be okay?

[ARKEYTECT]

Of course! Looking forward to it.

[ARKEYTECT]

Just curious, any update on the dereverberation model? Thanks!

[aask1357]

Sorry for the delay. I recently started training dereverberation + noise suppression model using a new dataset code. I am trying rts (https://arxiv.org/pdf/2204.08765) (https://arxiv.org/pdf/2505.01338) with t60_max=0.15 seconds and mixing reverberation into 90% of the training data.
I will keep updated under this issue page. Also, if you have any recommendations or tips for training dereverberation models, please feel free to share!

[aask1357]

I trained two versions for small size: t60_max=0.15 and 0.3 (https://github.com/aask1357/fastenhancer/releases/tag/ckpt-dereverb-v0.1.0). Currently I am not sure which version is better in terms of dereverb performance. Can you share your opinion about which one is better? Or any other opinion is welcomed. If you choose one, I will train the fastenhancer with other sizes too.

[ARKEYTECT]

I trained two versions for small size: t60_max=0.15 and 0.3 (https://github.com/aask1357/fastenhancer/releases/tag/ckpt-dereverb-v0.1.0). Currently I am not sure which version is better in terms of dereverb performance. Can you share your opinion about which one is better? Or any other opinion is welcomed. If you choose one, I will train the fastenhancer with other sizes too.

Apologize for late reply.. Thank you for training the model and spending the time! I listened to the output from both models, I think there are more distortions in t60_max=0.15 than in 0.3 version, but of course 0.15 version removed more reverb.
I think in terms of augmentation, you could perhaps look at the well-known DeepFilterNet3 augmentation code, other than that I don't have much suggestions..

Thanks!

[aask1357]

I uploaded other models too. I used rts0.15 because it matches the task of dereverb. If you have any thought about those models, feel free to share with me.