“2013-10-18”版本间的差异

2013年10月18日 (五) 09:30的版本

Data sharing

• LM count files still undelivered!

DNN progress

Sparse DNN

• Optimal Brain Damage(OBD). The initial test shows worse results in weight-cutting experiments compared with simple weight-based cutting.

Tencent exps

N/A

Noisy training

1. With 863 clean test, by adding car & white noise at various levels, obtained significant performance improvement.

• car noise test

[媒体文件:Noise-training-white-noise-test.png]

2. The test with both car & white noise benefits from the noisy training.

Continuous LM

• Lattice re-scoring toolkit is ready. However the toolkit is very slow for large lattices.
• Now checking the code to improve efficiency.

QA LM

Just started. Jobs to do:

1. use the QA-oriented word segment system
2. train the Q LM with the QA data