Example for speech enhancement

[JNaranjo-Alcazar]

Hi community! I was wondered if there is an example for speech enhancement. That is, given a noisy audio, the framework will return a cleaned audio with the speech.

Thanks in advance!

[mpariente]

Here is a demo, built by @JorisCos
https://asteroid-demo.herokuapp.com/

BTW, thanks for opening in discussion, this is the right place! 😃

[JNaranjo-Alcazar]

Thanks for the info!
Is the code used available somewhere? I mean, I can see an example about how to do a source separation, but I cannot see a speech enhancement one...
Thanks again!

[mpariente]

https://github.com/asteroid-team/asteroid_app

[JNaranjo-Alcazar]

Thank you very much partner! I definitely owe you one ;)

I'm still starting out in the speech domain... It's quite different from environmental audio haha

[mpariente]

By the way, you can also use the inference capabilities of the HF hub models, here for example.
We didn't enable it for all of them, but it should be simple to do.

[JNaranjo-Alcazar]

Ok, I'll have a look at the models that are available :)

[manu0586]

I just tried some pretrained models for SE
Basically you can perform speech enhancement in 2 lines (Big thanks at ASTEROID TEAM for that ) :

from asteroid.models import BaseModel
#SE-Model
if ID=='DPT':
    model= BaseModel.from_pretrained("JorisCos/DPTNet_Libri1Mix_enhsingle_16k")  #si-SDR : 14.7
if ID=='DPRNN':
    model=BaseModel.from_pretrained("JorisCos/DPRNNTasNet-ks2_Libri1Mix_enhsingle_16k") #si-SDR : 14.7
if ID=='DCCRN':
    model=BaseModel.from_pretrained("JorisCos/DCCRNet_Libri1Mix_enhsingle_16k")
if ID=='ConvTasNet':
    model=BaseModel.from_pretrained("JorisCos/ConvTasNet_Libri1Mix_enhsingle_16k") 

model.separate('test.wav')

The enhanced recording is saved in 'test_est1.wav'

[jonashaag]

You can also use the CLI to do inference, see the basic tutorial: http://colab.research.google.com/github/asteroid-team/asteroid/blob/master/notebooks/00_GettingStarted.ipynb

The tutorial is for separation but works the same for enhancement. Please tell if there is anything we could have explained better!