search

google / uis-rnn

This is the library for the Unbounded Interleaved-State Recurrent Neural Network (UIS-RNN) algorithm, corresponding to the paper Fully Supervised Speaker Diarization.
https://arxiv.org/abs/1810.04719
Apache License 2.0
1.55k stars 320 forks source link

Have you train the model on one database and test it on another database? #24

Closed wuqiangch closed 5 years ago

wuqiangch commented 5 years ago

In your paper, "We randomly partition the dataset into five subsets, and each time leave one subset for evaluation, and train UIS-RNN on the other four subsets. Then we combine the evaluation on five subsets and report the averaged DER." So the persons in traindatas also appeare in the testdata. Have you train the model on one database and test the model on another database? The persons in traindatas dont appeare in the testdata. I have done it ,but get the bad result.

wq2012 commented 5 years ago

Yes, we do have these results, as presented in Table 2.

You might be looking at an old version of the paper.

screenshot from 2019-01-08 10-34-15

Aurora11111 commented 5 years ago

@wuqiangch I have tested this project with my own datsets too,but get really bad result when I test my model on database which persons in traindatas dont appeare in the testdata.

wq2012 commented 5 years ago

@Aurora11111 Although our training of UIS-RNN is on a different domain than the test set in above highlighted experiments, our training of the speaker embedding model is based on a super large dataset covering 100K+ speakers, and from various microphones and acoustic environments.

If your cross-domain results are bad, very likely it's because your speaker embedding model is too weak, which is unrelated to the UIS-RNN model.