“VPR tasks”版本间的差异

2018年6月12日 (二) 02:54的最后版本

Task 1: 16k model noisy training

 i: large scale noisy training (7500 pnorm) 4.1-5.1. Done
 ii: model test: 5.1-5.7. Done.
 iii: full data training (7500+other dataset+more conditions). See (i). Done.
 OUTPUT: large reliable 16k model. Done.

Task 2: 8k model noisy training

 i: small scale verification: 5.1 -5.7. Done. 
 ii: large scale training: 5.14-6.1. @ Lantian. Done.
 OUTPUT: large reliable 8k model. Done.

Task 3: Full-info training

 i: programing and small test: 4.1-5.1. Done
 ii: full training for 7500 people: 5.1-6.1. Done.
 iii: adaptation for Ali dataset: 5.15-6.1. Done.
 iv: applying to 8k model. Done.
 OUTPUT: more reliable 16k and 8k model. Done.

Task 4: gated attention for backend scoring

 i: programing and small scale test: 6.15-6.22.
 ii: further large scale test: 6.23-6.30.
 OUTPUT: More accurate VPR backend.

Task 5: enrollment noisy traning/adaption

 i: Online noise-adding enrollment: 6.4-6.12
 ii: Offline data-augmentation enrollment: 6.12-6.22 --HOLD
 OUTPUT: More noise-robust VPR backend.

Task 6: deep speaker VAD

 i: SVM-VAD: 6.11-6.15.
 ii: DNN-VAD: 6.11-6.15.
 OUTPUT: More accurate VAD.