I'm working on a speaker recognition challenge.
I have already trained my model on the voxceleb2 dataset in triplet setup. Now, for the challenge, I have two sets.
enrollment (1 audio/subject) [IDs given]
test (random number of audios without IDs)
I need to report EER on the test.
Is it okay if I train my model on enrollment data too or it will be considered leakage/cheating while reporting EER?
Let me elaborate, In speaker verification/recognition, we have enrollment and trials data. In the experiment, let's say, I have 5 known subjects/speakers. I first record their speech and label them. This is my enrollment data.
Enrollment:
speaker 1 -> audio_1
speaker 2 -> audio_2
speaker 3 -> audio_3
speaker 4 -> audio_4
speaker 5 -> audio_5
Now, I take some random speech data from other sources + more audio data from the 5 speakers, this is my test data.
Test:
speaker 1 -> audio_11
speaker 2 -> audio_21
speaker 3 -> audio_31
speaker 4 -> audio_41
speaker 5 -> audio_51
random -> random_1
random -> random_2
Now, I will generate the trials from the test.
speaker 1, audio_11
speaker 3, audio_11
speaker 2, audio_21
speaker 4, random_1
speaker 1, random_2
I need to predict from the trials if audio_11 belongs to speaker 1 or not, audio_11 belongs to speaker 3 or not, audio_21 belongs to speaker 2 or not, etc. (based on audio similarity).
In my case, I'm segmenting the enrollment audio and training my model on them before making the predictions on the trial/test data.
Dear Zabir Al Nazi ,
Look the link, maybe useful.
Preprint Data augmentation enhanced speaker enrollment for text-depen...
Conference Paper Far-Field End-to-End Text-Dependent Speaker Verification Bas...
Regards,
Shafagat
Dear Zabir Al Nazi ,
I really did not get what was enrollment phase data for, but if you trained your model then I think it could be validation part of your model which could not be considered as part of your training, although it is for validating your model accuracy and error for seen data and help your model prevent overfitting, then if it is validation part then training on these data put your model into overfitting.
Thank you Amin Honarmandi Shandiz for your reply.
The enrollment phase data is actually the real speaker ID data on which I am going to perform prediction. It is not necessarily validation data, it would not overfit the model as the enrollment data and trials/test data are always different.
Let me elaborate, In speaker verification/recognition, we have enrollment and trials data. In the experiment, let's say, I have 5 known subjects/speakers. I first record their speech and label them. This is my enrollment data.
Enrollment:
speaker 1 -> audio_1
speaker 2 -> audio_2
speaker 3 -> audio_3
speaker 4 -> audio_4
speaker 5 -> audio_5
Now, I take some random speech data from other sources + more audio data from the 5 speakers, this is my test data.
Test:
speaker 1 -> audio_11
speaker 2 -> audio_21
speaker 3 -> audio_31
speaker 4 -> audio_41
speaker 5 -> audio_51
random -> random_1
random -> random_2
Now, I will generate the trials from the test.
speaker 1, audio_11
speaker 3, audio_11
speaker 2, audio_21
speaker 4, random_1
speaker 1, random_2
I need to predict from the trials if audio_11 belongs to speaker 1 or not, audio_11 belongs to speaker 3 or not, audio_21 belongs to speaker 2 or not, etc. (based on audio similarity).
In my case, I'm segmenting the enrollment audio, and training my model on them before making the predictions on the trial/test data.
your welcome Zabir Al Nazi ,
then you have your own audio and some random data which you know they are random,
and Am I right about bellow separation:?
(enrollment = speaker audios
test= speaker audios + random voices
training= speaker audios
)
and if these separation is right then you can train your model on your training data , why do you want to train it again on enrollment data?
- Badges
- Science topic