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

Question on the generative process of UIS-RNN #20

Closed 123456789077 closed 4 years ago

123456789077 commented 5 years ago

Hello， I haven't found a speaker change embodiment in the code. No training process was found for audio containing multiple speakers. In the pre-processing of training data, audio segments of the same individual is collected together, but how is the training process? Could you explain the generative process of uis-rnn？ Thank you in advance for your help.

taylorlu commented 5 years ago

1: The function sample_permuted_segments seems not work in utils.py, it generates always the same permutation. 2: And also, why need to collect the same cluster in resize_sequence, the overlapping window to calculate speaker embedding will not make sense from this way.

wq2012 commented 5 years ago

@taylorlu

  1. You mean the outputs in a single call are all identical sequences? Or do you mean when you call the function multiple times for the same input you always get same output? If latter, check if it is because you fixed the random seed of numpy.
  2. There's additional segment level aggregation logic after sliding windows. Segments are not overlapping. See section 2 of this paper.