2014-10-20

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         |  
  ------------------------------------------------------------

Sparse DNN

• Performance improvement found when pruned slightly
• Experiments show that
• Suggest to use TIMIT / AURORA 4 for training

RNN AM

• Initial test on WSJ , leads to out-memory.
• Using AURORA 4 short-sentence with a smaller number of targets.

Noise training

• First draft of the noisy training journal paper
• Paper Correction (Yinshi, Liuchao, Lin Yiye), be going.

Drop out & Rectification & convolutive network

• Drop out

• dataset:wsj, testset:eval92

       std |  dropout0.4 | dropout0.5 | dropout0.7 | dropout0.8
    ------------------------------------------------------------- 
       4.5 |     5.39    |    4.80    |   4.36     |    -      

• Test on noisy AURORA4 dataset

       std |  dropout0.4 | dropout0.5 | dropout0.7 | dropout0.8
    ------------------------------------------------------------- 
      6.05 |     -       |    -       |   -        |   -

• Continue the droptout on normal trained XEnt NNET , eg wsj. (+)
• Draft the dropout-DNN weight distribution. (+)

• Rectification

• Still NAN error, need to debug. (+)

• MaxOut (+)

• Convolutive network

• Test more configurations
• Yiye will work on CNN
• Reading CNN tutorial

Denoising & Farfield ASR

• ICASSP paper submitted.

VAD

• 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

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(+)
• Suggest to use Tencent training data(+)

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

 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        |       |       |        |

Scoring

• global scoring done.
• Pitch & rhythm done, need testing
• Harmonics program done, experiment to be done.

Confidence

• Reproduce the experiments on fisher dataset.
• Use the fisher DNN model to decode all-wsj dataset

Speaker ID

• Preparing GMM-based server.

Emotion detection

• Sinovoice is implementing the server

Text Processing

LM development

Domain specific LM

• lm based on baidu_hi and baiduzhidao is done, test on shujutang test set.
• weibo lm were training with pruning on counts(5,10,10,20,20),because it is too large. the ppl is twice as high than baidu_hi && baidu_zhidao.

• dongxu get good vocabulary from big data. Train 5-gram LM using Baiduzhidao_corpus(~30GB after preprocess) with new lexicon. There is a mistake when counted possiblity after merge.

tag LM

• use HIT's LTP tool to segment,pos and ner. the program is running(about 3 days) on baiduHi and baiduzhidao(total 365G)
• will use the small test set from xiaoxi for address-tag..
• now about more 1M address,will prune it using frequency.

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.

RNN LM

• rnn

get baseline on nbest rescore of wer.

• lstm+rnn

trained the RNN+LSTM lm on wsj_np_data about 200M. the neural net work is 100*100(lstm cell)*10000 with 100 classes. it cost about 200 minutes each epoch.

get baseline on nbest rescore of wer.

more detail on LSTM

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:

add the vsm and BM25 to improve the search. and the strategy of selecting the answer.

• spell check

get ngram tool and make a simple demo.

get domain word list and pingyin tool from huilan.

• new inter will install SEMPRE