2014-11-03
来自cslt Wiki
目录
Speech Processing
AM development
Environment
- buy two 760-GPUs.
- sale the old GPU.
Sparse DNN
- Performance improvement found when pruned slightly
- Experiments show that
- Waiting for result of AURORA 4
- HOLD
RNN AM
- Initial nnet seems no very well, need to be pre-trained or test lower learn-rate.
- For AURORA 4 1h/epoch, more than 200 epochs have done.
- Using AURORA 4 short-sentence with a smaller number of targets.
- Adjusting the learning rate.
- Trying toolkit of Microsoft.
Noise training
- Paper has been submitted.
Drop out & Rectification & convolutive network
- Drop out
- dataset:wsj, testset:eval92
std | dropout0.4 | dropout0.5 | dropout0.6 | dropout0.7 | dropout0.8
-------------------------------------------------------------------------
4.5 | 5.39 | 4.80 | 4.75 | 4.36 | 4.55
- Frame-accuarcy seems not consistent with WER.
- Using the train-data as cv, verify the learning ability of the model.
- AURORA4 dataset
(1) Train: train_clean
drop-retention/testcase(WER) | test_clean_wv1 | test_airport_wv1 | test_babble_wv1 | test_car_wv1
---------------------------------------------------------------------------------------------------------
std-baseline | 6.04 | 29.91 | 27.76 | 16.37
---------------------------------------------------------------------------------------------------------
dp-0.4 | 6.61 | 29.59 | 30.12 | 19.40
---------------------------------------------------------------------------------------------------------
dp-0.5 | 6.40 | 28.07 | 27.88 | 19.88
---------------------------------------------------------------------------------------------------------
dp-0.6 | 6.36 | 26.68 | 24.85 | 18.32
---------------------------------------------------------------------------------------------------------
dp-0.7 | 6.13 | 25.53 | 23.90 | 15.69
---------------------------------------------------------------------------------------------------------
dp-0.8 | 5.94 | 24.94 | 23.67 | 15.77
---------------------------------------------------------------------------------------------------------
dp-0.9 | 5.96 | 27.30 | 25.63 | 15.46
---------------------------------------------------------------------------------------------------------
(2) Train: train_nosiy
drop-retention/testcase(WER) | test_clean_wv1 | test_airport_wv1 | test_babble_wv1 | test_car_wv1
---------------------------------------------------------------------------------------------------------
std-baseline | 9.60 | 11.41 | 11.63 | 8.64
---------------------------------------------------------------------------------------------------------
dp-0.3 | 12.91 | 16.55 | 15.37 | 12.60
---------------------------------------------------------------------------------------------------------
dp-0.4 | 11.48 | 14.43 | 13.23 | 11.04
---------------------------------------------------------------------------------------------------------
dp-0.5 | 10.53 | 13.00 | 12.89 | 10.24
---------------------------------------------------------------------------------------------------------
dp-0.6 | 10.02 | 12.32 | 11.81 | 9.29
---------------------------------------------------------------------------------------------------------
dp-0.7 | 9.65 | 12.01 | 12.09 | 8.89
---------------------------------------------------------------------------------------------------------
dp-0.8 | 9.79 | 12.01 | 11.77 | 8.91
---------------------------------------------------------------------------------------------------------
dp-1.0 | 9.94 | 11.33 | 12.05 | 8.32
---------------------------------------------------------------------------------------------------------
baseline_dp0.4_lr0.008 | 9.52 | 12.01 | 11.75 | 9.44
---------------------------------------------------------------------------------------------------------
baseline_dp0.4_lr0.0001 | 9.92 | 14.22 | 13.59 | 10.24
---------------------------------------------------------------------------------------------------------
baseline_dp0.4_lr0.00001 | 9.06 | 13.27 | 13.14 | 9.33
---------------------------------------------------------------------------------------------------------
baseline_dp0.8_lr0.008 | 9.16 | 11.23 | 11.42 | 8.49
---------------------------------------------------------------------------------------------------------
baseline_dp0.8_lr0.0001 | 9.22 | 11.52 | 11.77 | 8.82
---------------------------------------------------------------------------------------------------------
baseline_dp0.8_lr0.00001 | 9.12 | 11.27 | 11.65 | 8.68
---------------------------------------------------------------------------------------------------------
dp-0.4_follow-std-lr | 11.33 | 14.60 | 13.50 | 10.95
---------------------------------------------------------------------------------------------------------
dp-0.8_follow-std-lr | 9.77 | 12.01 | 11.79 | 8.93
---------------------------------------------------------------------------------------------------------
dp-0.4_4-2048 | 11.69 | 16.13 | 14.24 | 11.98
---------------------------------------------------------------------------------------------------------
dp-0.8_4-2048 | 9.46 | 11.60 | 11.98 | 8.78
---------------------------------------------------------------------------------------------------------
- Test baseline of 2048 dimension.
- Test with AURORA4 of 7000 (clean + noisy).
- Follow the standard DNN training learn-rate to avoid the different learn-rate changing time of various DNN training. Similar performance is obtained.
- Find and test unknown noise test-data.(+)
- Have done the droptout on normal trained XEnt NNET , eg wsj(learn-rate:1e-4/1e-5).
- Draft the dropout-DNN weight distribution. (++)
- Rectification
- Still NAN error, need to debug.
1) AURORA 4 -15h
(1) Train: train_clean
learn-rate/testcase(WER) | test_clean_wv1 | test_airport_wv1 | test_babble_wv1 | test_car_wv1
---------------------------------------------------------------------------------------------------------
std-baseline | 6.04 | 29.91 | 27.76 | 16.37
---------------------------------------------------------------------------------------------------------
lr0.00001 | 8.30 | 43.85 | 46.42 | 29.80
---------------------------------------------------------------------------------------------------------
lr0.0001 | 6.57 | 31.11 | 30.65 | 19.65
---------------------------------------------------------------------------------------------------------
lr0.0006 | 6.19 | 29.23 | 28.45 | 17.31
---------------------------------------------------------------------------------------------------------
lr0.0008 | 6.17 | 28.10 | 27.46 | 14.97
---------------------------------------------------------------------------------------------------------
lr0.001 | 6.28 | 30.01 | 30.26 | 20.81
---------------------------------------------------------------------------------------------------------
lr0.003 | 6.44 | 32.01 | 32.24 | 17.82
---------------------------------------------------------------------------------------------------------
lr0.005 | 6.47 | 33.49 | 34.75 | 18.15
---------------------------------------------------------------------------------------------------------
lr0.007 | 6.72 | 35.85 | 39.72 | 18.03
---------------------------------------------------------------------------------------------------------
lr-0.001_l1-0.001 | 83.19 | 98.57 | 98.84 | 97.77
---------------------------------------------------------------------------------------------------------
lr-0.001_l1-0.0001 | 7.58 | 32.94 | 34.29 | 23.42
---------------------------------------------------------------------------------------------------------
lr-0.001_l1-0.00001 | 6.21 | 29.15 | 28.24 | 19.50
---------------------------------------------------------------------------------------------------------
lr-0.001_l1-0.000001 | 6.30 | 31.91 | 29.23 | 21.52
---------------------------------------------------------------------------------------------------------
- Train rectifier on baseline.
- Combine drop out and rectifier.
- Change the learn-rate in the middle of the training, Modify the train_nnet.sh script(Liu Chao).
- MaxOut (+)
- 6min/epoch, can't use high lr.
- P-norm
- Convolutive network (+)
ORG_DNN.4-1200 WER: 4.50
| WER | hid-layers | hid-dim | delta-order | splice | lda-dim | learn-rate | cnn_init_opts ------------------------------------------------------------------------------------------------------------------ cnn_std_baseline | 4.86 | 5 | 1200 | 0 | 5 | 198 | 0.008 | "--patch-dim1 7"
- AURORA 4
- READ paper
Denoising & Farfield ASR
- ICASSP paper submitted.
- HOLD
VAD
- Spike detection and removal.
- Sum all sil-pdf as the silence posterior probability.
- Program done, to tune the threshold
- rearrange the ending point of the detected speech
Speech rate training
- Seems ROS model is superior to the normal one with faster speech
- Suggest to extract speech data of different ROS, construct a new test set(+)
- Tencent training data with 100h
low resource language AM training
- Use Chinese NN as initial NN, change the last layer
- Various the used Chinese trained DNN layer numbers.
- feature_transform = 6000h_transform + 6000_N*hidden-layers
- Various the used Chinese trained DNN layer numbers.
nnet.init = random (4-N)*hidden-layers + output-layer | N / learn_rate | 0.008 | 0.001 | 0.0001 | | baseline | 17.00(14*2h) | | | | 4 | 17.75(9*0.6h) | 18.64 | | | 3 | 16.85 | | | | 2 | 16.69 | | | | 1 | 16.87 | | | | 0 | 16.88 | | |
mpe: | Nnet-structure | WER | | baseline | 16.12 | | 4-0-1 | | | 4-1-1 | 16.10 | | 4-2-1 | 15.66 | | 3-1-1 | 16.10 | | 3-2-1 | 15.64 | | 2-2-1 | 15.73 | | 1-3-1 | 15.91 | | 0-4-1 | |
remark: 4-0-1 means 4 hidden-layers from 6000h_CN, 0 hidden-layer from random generation, 1 output-layer.
- feature_transform = uyghur_transform + 6000_N*hidden-layers
nnet.init = random (4-N)*hidden-layers + output-layer Note: This is reproduced Yinshi's experiment | N / learn_rate | 0.008 | 0.001 | 0.0001 | | baseline | 17.00 | | | | 4 | 28.23 | 30.72 | 37.32 | | 3 | 22.40 | | | | 2 | 19.76 | | | | 1 | 17.41 | | | | 0 | | | |
- feature_transform = 6000_transform + 6000_N*hidden-layers
nnet.init = uyghur (4-N)*hidden-layers + output-layer | N / learn_rate | 0.008 | 0.001 | 0.0001 | | baseline | 17.00 | | | | 4 | 17.80 | 18.55 | 21.06 | | 3 | 16.89 | 17.64 | | | 2 | | | | | 1 | | | | | 0 | | | |
- sub word unit language model is ready. on testing.
Scoring
- Timber Comparison on testing
Confidence
- Reproduce the experiments on fisher dataset.
- Use the fisher DNN model to decode all-wsj dataset
- preparing scoring for puqiang data
Speaker ID
- Preparing GMM-based server.
- EER ~ 11.2% (GMM-based system)
- test different number of components; fast i-vector computing
Language ID
- GMM-based language is ready.
- Delivered to Jietong
Emotion detection
- Sinovoice is implementing the server
Text Processing
LM development
Domain specific LM
- domain lm
- weibo lm with pruning 0 10 10 20 20 testing done. weibo lm with pruning 0 10 8 8 8 under testing. weibo lm without pruning 4/8 done.
- merger weibo、baiduhi and baiduzhidao lm and test (this week)
- confirm the size of alpa with xiaomin for business application.(like e-13)
- get the general test data from miaomin .this test set may get from online.
- new dict.
- train lm on baiduhi, baiduzhida with new 150k dict and test (this week)
- new toolkit:find method to update the new dict. can get new wordlist from sougou and get word information from baidu.(two week)
tag LM
- set new test
- fix the bug
- record test set and test the unknown address (this week)
RNN LM
- rnn
- RNNLM=>ALPA make a report
- test RNNLM on Chinese data from jietong-data
- check the rnnlm code.
- lstm+rnn
- check the lstm-rnnlm code
Word2Vector
W2V based doc classification
- Initial results variable Bayesian GMM obtained. Performance is not as good as the conventional GMM.
- Non-linear inter-language transform: English-Spanish-Czch: wv model training done, transform model on investigation
- SSA-based local linear mapping still on running.
- k-means classes change to 2.
- Knowledge vector started
- format the data
- yuanbin will continue this work with help of xingchao.
- Character to word conversion
- prepare the task: word similarity
- prepare the dict.
- Google word vector train
- some ideal will discuss on weekly report.
Translation
- v4.0 demo released
- cut the dict and use new segment-tool
QA
- lucene Optimization
- rewrite the method to select the 50 standard question not same template.(this week)
- test the boost keyword weight and extract the synonyms word.(this week)
- check the word segment for template.(this week)
- min-segment method improve the accuracy.(0.61->0.66)
- check the query method for getting lucene information and to rewrite the score method like the idf value.
- test
- test the different idf vale from baidu sougou in fuzzymatch.(this week)
- need to check the other 10% error.(this week)
- spell check
- simple demo done.
- new inter will install SEMPRE
