2014-10-27
来自cslt Wiki
Speech Processing
AM development
Contour
- NAN problem
- nan recurrence
------------------------------------------------------------ grid/atr. | Reproducible | add. ------------------------------------------------------------ grid-10 | yes | ------------------------------------------------------------ grid-12 | no | "nan" in different position ------------------------------------------------------------ grid-14 | yes | ------------------------------------------------------------
- buy 760-GPU
Sparse DNN
- Performance improvement found when pruned slightly
- Experiments show that
- Suggest to use TIMIT / AURORA 4 for training
- HOLD
RNN AM
- Initial nnet seems no very well, need to be pre-trained or test lower learn-rate.
- For AURORA4 1h/epoch, 100 epochs done.
- Using AURORA 4 short-sentence with a smaller number of targets.
Noise training
- First draft of the noisy training journal paper.
- Second version released.
- Paper Correction (Yinshi, Liuchao, Lin Yiye), be going.
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
---------------------------------------------------------------------------------------------------------
- Losing important features, enlarge the hidden-layer dim to 2048.
- Follow the standard dnn training learn-rate to avoid the different learn-rate changing time of various DNN training.
- Test out of known noise test-data.
- Continue 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) AURORA4 -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
---------------------------------------------------------------------------------------------------------
lr0.001 | 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
---------------------------------------------------------------------------------------------------------
- Change the learn-rate in the middle of the training, Modify the train_nnet.sh script(Liu Chao).
- Using maximum learning-rate.
- MaxOut (++)
- Convolutive network (+)
- Test more configurations
Denoising & Farfield ASR
- ICASSP paper submitted.
- HOLD
VAD
- Spike detection and removal.
- Add more silence tag "#" in pure-silence utterance text(train).
- xEntropy model be training
- need to test baseline.
- 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 done
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 | | |
- 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
- Harmonics program done, experiment to be done.
- Initial experiment shows more timber data are required
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
Emotion detection
- Sinovoice is implementing the server
Text Processing
LM development
Domain specific LM
- domain lm
- am:1400h(2.0.b) .result: xiaomi-29.43%,baiduzhidao-43.46%,baiduHi-30.02%, test-set:8ksentence(16k=>8k)
- need to check the xiaomin-lm method and result.
- new dict.
- weibo-data : Tencent-segment and count. get 16k words to segment again.
- new toolkit:find method to update the new dict. can get new wordlist from sougou and get word information from baidu.
tag LM
- set new test
- 1k address from dianxin. prepare to test.
- insert the new unknown-address to test set.
- record test set 15-sentence/person on dianxin txt.
RNN LM
- rnn
- RNNLM=>ALPA
- train RNNLM on Chinese data from jietong-data
- lstm+rnn
- wer:6.2%(4-epoch).need to check the problem.
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
- v3.0 demo released
- still slow
- re-segment the word using new dictionary.will use the tencent-dic about 11w.
- check new data.
QA
- search method:
- test the lucene method
- analysis the test result
- add IDF to test
- spell check
- get ngram tool and make a simple demo.
- get domain word list and pingyin tool from huilan.
- new inter will install SEMPRE
