How to generate enrollment data for testing CTDNN model?

[bjfu-ai-institute]

How to generate enrollment data for testing CTDNN model?

[vzxxbacq]

Hi @gamesterrishi ,
In morden speaker recognition field, we want to build a model apply for every person even if the person hasn't shown up in the train dataset. One solution is calculating the subspace vector that we use a model to transfer the input feature to speaker vectors. ( such as i-vector, d-vector ), then we use these vectors to recognize each person. So, if we want to recognize a person, we need at least one audio of him/her to calculate the speaker vector, we call this step enrollment. Then in test step, we get the speaker vector from a unknow audio and use enrolled vectors to determine who spoke, if the speaker hasn't enrolled, we call him/her intruder.
So, if you don't want to do some experiment like intruder testing, you just need to cut a audio of each person from test dataset. And one thing you should know, don't copy the test dataset, because obviously same audio get same vectors, we want our model can extract similar vectors from same person's different audio.

[gamesterrishi]

Hi @vzxxbacq, thanks for the information. I understand that we have to go through this approach for the process of speaker verification. However I wanted to use this model for the purpose of speaker recognition - recognizing the speaker id of the person who speaks.
One of the approach to do this according to my understanding:

1. For each speaker that the model has been trained on, create 400 dimensional d vectors specific for all speakers and store them in an array
2. On receiving the audio of the unknown speaker whose identity we wish to find, get features extracted from that audio file and get predict the d vector for this audio file from the second last layer and do a cosine similarity of this d vector with d vectors of all speakers --> highest similarity will give us the speaker id. Can you please let me know if that is right?

Also you have added some good backend to your project - I am not able to understand where and how can I make use of it. Thanks for your help.

[vzxxbacq]

@gamesterrishi , Sorry for my late reply, your method is correct and we can also use other classification model to calculate the score, such as SVM, LDA and PLDA. I was using Sidekit to do backend jobs and now I'm trying to implement PLDA and LDA by myself but I'm not sure if it will work properly before I test it with true dataset.