Attention_Backend_for_ASV

Attention Backend for Aotumatic Speaker Verification with Multiple Enrollment Utterances

It contains the official implementation of the paper Attention Back-end for Automatic Speaker Verification with Multiple Enrollment Utterances

requirements

1. Kaldi. And set kaldi path in path.sh and run.sh according to the instruction in these files.
2. Pytorch >= 1.0.0
3. Numpy

data

You can download the data from this link

Password: e2de.

It contains x-vectors extracted by the script of cnceleb example in Kaldi (train_xv, enroll_xv, eval_xv)

results

Note: This code only for X-Vectors extracted from TDNN.

to do

• TD-ASV (RedDots)

• Change score method from cosine similarity to PLDA-like score

• Breakdown results per domain (genre) in CN-Celeb

reference

siamese-triplet