Causal Diffusion Models for Generalized Speech Enhancement
On this website, we present listening examples for the work "Causal Diffusion Models for Generalized Speech Enhancement". For both causal and non-causal settings, we compare the generalized model trained on the entire dataset (denoted as "All") with specialized models only trained on the respective corruption set ("Noise", "Reverb", "Clipping", "PacketLoss", and "BWR+Codecs").
Additive background noise
Method |
Training set |
File 1 |
File 2 |
File 3 |
File 4 |
---|---|---|---|---|---|
Clean | |||||
Unprocessed | |||||
Causal | Noise | ||||
Causal | All | ||||
Non-causal | Noise | ||||
Non-causal | All |
Reverberation
Method |
Training set |
File 1 |
File 2 |
File 3 |
File 4 |
---|---|---|---|---|---|
Clean | |||||
Unprocessed | |||||
Causal | Reverb | ||||
Causal | All | ||||
Non-causal | Reverb | ||||
Non-causal | All |
Clipping
Method |
Training set |
File 1 |
File 2 |
File 3 |
File 4 |
---|---|---|---|---|---|
Clean | |||||
Unprocessed | |||||
Causal | Clipping | ||||
Causal | All | ||||
Non-causal | Clipping | ||||
Non-causal | All |
Packet loss
Method |
Training set |
File 1 |
File 2 |
File 3 |
File 4 |
---|---|---|---|---|---|
Clean | |||||
Unprocessed | |||||
Causal | PacketLoss | ||||
Causal | All | ||||
Non-causal | PacketLoss | ||||
Non-causal | All |
Bandwidth reduction and codecs
Method |
Training set |
File 1 |
File 2 |
File 3 |
File 4 |
---|---|---|---|---|---|
Clean | |||||
Unprocessed | |||||
Causal | BWR+Codecs | ||||
Causal | All | ||||
Non-causal | BWR+Codecs | ||||
Non-causal | All |