Lantian Li 2015-10-19

Weekly Summary

1. Go on my deep speaker embedding tasks:

1). knowledge transfer for i-vector -- working with Zhiyuan Zhang.

Finish modifying the nnet-training objective function (nnet-train-perutt-persoft and nnet-train-perutt-persoft-average).

Finish DNN training process.

2). Start my new task: metric learning using deep neural networks.

Using 'Wangd dataset' to re-compare the non-linear(two layers NN) and linear metric learning.

On processing.

2. Re-embedding i-vector with lower dimension 200 and lda 100.

Note that for the process of T-matrix training, Num_threads should be set 1.

3. Complete my thesis proposal.

4. Write paper on 'Discriminative Score Feature Selection for Speaker Verification'.

Next Week

1. Go on the task 1.

Decoding 1) and 2).