Ctdnn approach dount

[bjfu-ai-institute]

Ctdnn approach dount

[vzxxbacq]

Hi Mihir, I'm glad to know my work is useful to you~

In this paper, authors use 40-d Fbank feature and concatenate 4 frames' feature before and after each frame. It's easy to implement, you just need to write code like this feature[index] = feature[index-4 : index+5]. So, the features' shape is (9, 40) and the 1 represents channel.

If you want use MFCC feature, you can use 40-d MFCC feature or change 40 to 39, but you need also to modify the conv_kernel and stride.

[mihirp1998]

Thanks a lot

[vzxxbacq]

@mihirp1998 I saw the email you sent me, but I didn't see the reply here, so weird.

I glad to know my repo is useful to you, and I also hope you would like to pull request after you implement them. If you have any questions about my code, feel free to contact me.

About the work on SV. Recently I am studying the effect of data augmentation (e.g. add noise) on the speaker recognition model. Although I have not made much progress now, I think this is feasible.

[mihirp1998]

nice to hear that. I'm currently working with a team of 3 .. we r working on noise cancellation techniques rather, like ltsd and few other. We will soon commit it to ur repo. Haven't really looked into noise augmentation. Can u share me some relevant links about it.

[vzxxbacq]

@mihirp1998
Thanks for your contribution so much.
This is link to data aug paper.