[Question]Performance degrade for different win size

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]Performance degrade for different win size #38

Closed FengLeee closed 5 years ago

FengLeee commented 5 years ago

Describe the question

i train embedding lstm with variable length batch size, about [24，160]，then i get d-vector with win size 160frame in this model, and use the d-vector train uis-rnn, the uis-rnn predict result is good(90%)。but when i use win size 40frame get the d-vector and train uis-rnn the uis-rnn predict result is bad (40%)。what the reason cause this problem？

wq2012 commented 5 years ago

Please clarify your setup:

[24, 160]: What's the unit here? Frame or ms? If frame, what's your frame size and step?
40frame: Again, what's your frame size and step?
win: Does your window overlap? If yes, how much?
Which training and testing dataset are you using for uis-rnn? Which dataset are you using for d-vector training?