From cslt Wiki
- quota on /nfs/disk this Saturday
- release management should be started: Zhiyong
- Blaster 0.1 & vivian 0.0 system release
- Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test. Overfitting?
- Multi GPU training: Error encountered
- Multilanguage training
- Investigating LOUDS FST.
- CLG embedded decoder plus online compiler.
- DNN-GMM co-training
- GA-based block sparsity
- Found a paper in 2000 with similar ideas.
- Try to get a student working on high performance computing to do the optimization
- More experiments with no-noise
- More experiments with additional noise types
AMR compression re-training
- 1700h MPE adaptation done
- 1700h stream mode adaptation runs into MPE4 done
- Stream model is better than non-stream wave
- GFBank Sinovoice test on 100h MPE
- Tencent 100h MPE training done
- all phone strategy baseline done
- Testing on Mandarin & English individually
Denoising & Farfield ASR
- re-Recording done
- Prepare to construct the baseline
- Code ready, migrate to the VAD code framework
- g-score based on MLP is done
- t-score based on linear regression improves the performance
Word to Vector
- Dimension of low space varies from 10-100
- 8-thread word vector generation is 3 times faster than the LDA.
- Character-based NNLM (6700 chars, 7gram), 500M data training done.
- Non-boundary char LM is better than boundary char LM
- Investigate MS RNN LM training
- Word-based FST 1-2 seconds with 1600 patterns. Huilan's implementation <1 second. ????
- Char-FST Implementation is done. Not so effective.
- Investigate determinization of G embedding