Lantian Li 15-07-21

来自cslt Wiki
2015年7月23日 (四) 01:04Lilt讨论 | 贡献的版本

(差异) ←上一版本 | 最后版本 (差异) | 下一版本→ (差异)
跳转至: 导航搜索

Weekly Summary

1. Prepare for three deep speaker embedding tasks:

1). large-scale deep speaker vector framework: -- hold.

2). knowledege transfer for i-vector: --hold.

3). derive binary i-vectors using Hamming distance learning: results as shown in CVSS 373.

Results shows that the performance of LSH for both i-vector and lda are inferior to the baselines.

Besides,the performance of Variable-blockTraining for lda outperforms the baseline under condition 1 and 3 on SRE08.

While the Variable-blockTraining for i-vector is uneffective.

2. SUSR experiments: find and organize the database and the baseline has been done.

EER of Cosine distance: 39.11%, EER of LDA: 20.80%. And the GMM-UBM from the Chenhao Paper is 29.78%.

Prepare the experiments on speech unit classes.


Next Week

1. Go on the task 1 and attempt to complete 3).

2. Complete the task2.

3.Help Prof.Thomas write the patent: Recording replay detection.