“2013-04-26”版本间的差异
来自cslt Wiki
| 第60行： | 第60行： | ||
===L-1 sparse initial training=== | ===L-1 sparse initial training=== | ||
| − | : | + | : L-1=1e-5, starting from 6th iteration, converged with another 3 iterations. The performance is generally worse than the case where l1=0, except one test suite. |
| + | : L-1=1e-6, the same results obtained, means le-6 is too small to be effective. | ||
| + | : L-1=1e-4, start from the first iteration. crashed. Need more investigation. | ||
==Kaldi/HTK merge== | ==Kaldi/HTK merge== | ||
:* HTK2Kaldi: hold. | :* HTK2Kaldi: hold. | ||
| − | :* Kaldi2HTK: done with implementation. | + | :* Kaldi2HTK: done with implementation. A bug fixed. gConst was computed in a wrong way. The current HDecode result is 14.9%; The tencent model is 11%; Kaldi decoder 7%. |
| + | :* Possibly the SP model issue, due to the complicated structure of silence in Kaldi. | ||
==Embedded progress== | ==Embedded progress== | ||
| − | :* PocketSphinx migration done. | + | :* PocketSphinx migration done, using PocketSphinx default Chinese model. After migrating to the smart phone, the test shows that the decoding is very slow. RT=7.0. |
:* QA LM training, done. | :* QA LM training, done. | ||
| + | |||
| + | :* Next substitute the LM with JSGF grammar involving 1000 words. Finish the initial test. | ||
| + | :* Need to train a new AM. | ||
2013年4月26日 (五) 06:32的版本
目录
Data sharing
- LM count files are still in transfering.
DNN progress
400 hour DNN training
| Test Set | Tencent Baseline | bMMI | fMMI | BN(with fMMI) | Hybrid |
|---|---|---|---|---|---|
| 1900 | 8.4 | 7.65 | 7.35 | 6.57 | 7.27 |
| 2044 | 22.4 | 24.44 | 24.03 | 21.77 | 20.24 |
| online1 | 35.6 | 34.66 | 34.33 | 31.44 | 30.53 |
| online2 | 29.6 | 27.23 | 26.80 | 24.10 | 23.89 |
| map | 24.5 | 27.54 | 27.69 | 23.79 | 22.46 |
| notepad | 16 | 19.81 | 21.75 | 15.81 | 12.74 |
| general | 36 | 38.52 | 38.90 | 33.61 | 31.55 |
| speedup | 26.8 | 27.88 | 26.81 | 22.82 | 22.00 |
- Tencent baseline is with 700h online data+ 700h 863 data, HLDA+MPE, 88k lexicon
- Our results are with 400 hour AM, 88k LM. ML+bMMI.
- The CSLT structure: 300*[1200*1200*1200*40*1200]*4850.
- The CSLT feature: MFCC+delta MFCC
- compare with the traditional structure 300*[1200*1200*1200*1200*1200]*4850.
Tencent test result
- AM: 70h training data
- LM: 88k LM
- Test case: general
| Feature | GMM | GMM-bMMI | DNN | DNN-MMI |
|---|---|---|---|---|
| PLP(-5,+5) [Eryu] | 47 | 38.4 | 26.5 | 23.8 |
| PLP+LDA+MLLT(-5,+5)[Jingbo] | 47 | - | 34 |
- Tencent NN structure:
- 300*[1200*1200*1200*1200]*1700, #param=700k
- 300*[1007*1007*1007*1007]*3xxx #param=700k
- CSLT reproduce phone-clustered based NN
- CSLT investigate performance of different epochs.
GPU & CPU merge
- Investigate the possibility to merge GPU and CPU code.
- CUDA code merged to CPU.
L-1 sparse initial training
- L-1=1e-5, starting from 6th iteration, converged with another 3 iterations. The performance is generally worse than the case where l1=0, except one test suite.
- L-1=1e-6, the same results obtained, means le-6 is too small to be effective.
- L-1=1e-4, start from the first iteration. crashed. Need more investigation.
Kaldi/HTK merge
- HTK2Kaldi: hold.
- Kaldi2HTK: done with implementation. A bug fixed. gConst was computed in a wrong way. The current HDecode result is 14.9%; The tencent model is 11%; Kaldi decoder 7%.
- Possibly the SP model issue, due to the complicated structure of silence in Kaldi.
Embedded progress
- PocketSphinx migration done, using PocketSphinx default Chinese model. After migrating to the smart phone, the test shows that the decoding is very slow. RT=7.0.
- QA LM training, done.
- Next substitute the LM with JSGF grammar involving 1000 words. Finish the initial test.
- Need to train a new AM.
