Different input size between training d-vector modle and creating dvector for uis-rnn

[wuqiangch]

when training d-vector model ,the input size is batchsizex180x40,but in creating dvecor feature for uis-rnn ,the input size of the d-vector model is batchsizex24x40.Does it matter?

[HarryVolek]

It does not matter in terms of the code working, but it does matter for performance of the https://github.com/google/uis-rnn clustering algorithm.

The solution mentioned in https://arxiv.org/abs/1810.04719 is to train the model with a variable amount of input frames picked between 24 and 160 from a uniform distribution (vary batch-by-batch). I have yet to implement and test this.

With that being said I was able to obtain OK results using https://github.com/google/uis-rnn with the code as-is.

[wuqiangch]

@HarryVolek Thanks! Have you release the code to train the d-vector model with a variable amount of input frames picked between 24 and 160 from a uniform distribution (vary batch-by-batch)? I can't find the code in data_load.py.

[HarryVolek]

Not at this moment.