search

astorfi / 3D-convolutional-speaker-recognition

:speaker: Deep Learning & 3D Convolutional Neural Networks for Speaker Verification
Apache License 2.0
780 stars 274 forks source link

validate accuracy on development dataset #34

Closed deyituo closed 6 years ago

deyituo commented 6 years ago

Hi astorfi, Thanks for your great job. These days I am running your code on my dataset but I found it the validate accuracy is low in my experiments. I have no idea if there is something wrong. What's the validate accuracy of your experiments when the network is converged?

SpongebBob commented 6 years ago

I have the same question。Can you leave me a wechat or QQ？ Any help will thank a lot。

astorfi commented 6 years ago

How to do you perform the experiments? You mentioned, "What's the validate accuracy of your experiments when the network is converged". The paper has three phases: Development, Enrollment, and Evaluation. Are you performing them correctly in order?

astorfi commented 6 years ago

You mentioned, "validate accuracy on development dataset". The architecture is not designed to have good validation accuracy for the development dataset by simply using the softmax. It is only for creating the background model.