“ASR:2014-12-01”版本间的差异

2014年12月1日 (一) 07:49的版本

Speech Processing

AM development

Environment

• Already buy 3 760GPU
• grid-9/12 760GPU crashed again; grid-11 shutdown automatically.
• Change 760gpu card of grid-12 and grid-14(+).

Sparse DNN

• Performance improvement found when pruned slightly
• need retraining for unpruned one; training loss
• details at http://liuc.cslt.org/pages/sparse.html
• HOLD

RNN AM

• Initial nnet seems not very well, need to be pre-trained or test lower learn-rate.
• For AURORA 4 1h/epoch, model train done.
• Using AURORA 4 short-sentence with a smaller number of targets.(+)
• Adjusting the learning rate.(+)
• Trying toolkit of Microsoft.(+)
• details at http://liuc.cslt.org/pages/rnnam.html
• Reading papers

A new nnet training scheduler

• Initial code done. No better than original one considering of taking much more iterations.
• details at http://liuc.cslt.org/pages/nnet-sched.html

Drop out & Rectification & convolutive network

• Drop out(+)

• AURORA4 dataset

• Use different proportion of noise data to investigate the effect of xEnt and mpe and dropout
  • Problem 1) The effect of dropout in different noise proportion;

          2) The effect of MPE in different noise proportion;
          3) The effect of MPE+dropout in different noise proportion.

• • http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?step=view_request&cvssid=261

• • Find and test unknown noise test-data.(++)
  • Have done the droptout on normal trained XEnt NNET , eg wsj(learn-rate:1e-4/1e-5). Seems small learn-rate get the balance of accuracy and train-time.
  • Debug the low cv frame-accuracy

• MaxOut(+)

• pretraining based maxout
  • Select units in Groupsize interval, but need low learn-rate
  • Force accept the first iteration. Jump out from the local-minimum

• SoftMaxout

• P-norm

• Need to solve the too small learning-rate problem
  • Add one normalization layer after the pnorm-layer

• Convolutive network (+)

• AURORA 4

nonlda                 | %WER      |Dnn l-u    | pool size-step| cnn dim-step-num                | cnn_init_opts

cnn_std                | 5.73      | 4 - 1200  | 3 - 3         | 8-1-128 512-128-256             |--patch-dim1 8 
                       |           |           |               |                                 |--input_dim~patch-dim1

cnn_cnnunit_384        | 5.85      | 4 - 1200  | 3 - 3         | 8-1-128 512-128-384             |--patch-dim1 8
                       |           |           |               |                                 |--num-filters2 384     

cnn_patchdim1_5        | 5.92      | 4 - 1200  | 3 - 3         | 5-1-128 512-128-256             |--patch-dim1 5

cnn_patchdim1_11       | 6.05      | 4 - 1200  | 3 - 3         | 11-1-128 512-128-256            |--patch-dim1 11

cnn_delta_1            | 5.98      | 4 - 1200  | 3 - 3         | 8-1-128 512-128-256             |--patch-dim1 8

cnn_delta_2            | 6.05      | 4 - 1200  | 3 - 3         | 8-1-128 512-128-256             |--patch-dim1 8

cnn_layer_3            | 6.00      | 4 - 1200  | 3 - 3 3 - 1   | 8-1-128 512-128-256 768-256-512 |--patch-dim1 8

cnn_layer_3_2          | 5.85      | 4 - 1200  | 3 - 3 2 - 2   | 8-1-128 512-128-256 768-256-512 |--patch-dim1 8

cnn_layer_3_3          | 5.73      | 4 - 1200  | 3 - 3 2 - 2   | 8-1-128 512-128-256 512-256-512 |--patch-dim1 8

cnn_layer_3_4          | 5.96      | 4 - 1200  | 3 - 3 2 - 2   | 8-1-128 512-128-256 256-256-512 |--patch-dim1 8

DAE(Deep Atuo-Encode)

 (1) train_clean
   drop-retention/testcase(WER)| test_clean_wv1  | test_airport_wv1 | test_babble_wv1 | test_car_wv1 
  ---------------------------------------------------------------------------------------------------------
      std-xEnt-sigmoid-baseline| 6.04            |    29.91         |   27.76         | 16.37
  ---------------------------------------------------------------------------------------------------------
      std+dae_cmvn_noFT_2-1200 | 7.10            |    15.33         |   16.58         | 9.23
  ---------------------------------------------------------------------------------------------------------
   std+dae_cmvn_splice5_2-100  | 8.19            |    15.21         |   15.25         | 9.31
  ---------------------------------------------------------------------------------------------------------

Denoising & Farfield ASR

• ICASSP paper submitted.
• HOLD

VAD

• Frame energy feature extraction, done
• Harmonics and Teager energy features being investigation (++)
• Previous results to be organized for a paper
• MPE model VAD ,good performance observed.

Speech rate training

• Data ready on tencent set; some errors on speech rate dependent model
• Retrain new model(+)

Scoring

• Timber Comparison done.
• harmonics based timber comparison: frequency based feature is better
• GMM based timber comparison is done. Similar to speaker recognition
• TODO: Code checkin and technique report

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 ~ 4% (GMM-based system)--Text independent
• EER ~ 6%(1s) / 0.5%(5s) (GMM-based system)--Text dependent
• test different number of components; fast i-vector computing

Language ID

• GMM-based language is ready.
• Delivered to Jietong
• Prepare the test-case

Voice Conversion

• Yiye is reading materials

Text Processing

LM development

Domain specific LM

• domain lm(need to discuss with xiaoxi)

• embedded language model(this week)
• train some more LMs with Zhenlong (dianzishu sogou bbs chosen)("need result").
• keep on training sogou2T lm(14/16 on 3rd iteration).(this week)

• new dict.

• handover of this work to hanzhenglong, give a simple docuemnt(this week)

tag LM

• different weight 2014-Nov-23,Monday

• conclusion:

 1. compare experiment 3  with experiment 5:
   same jsgf file, but the  tag number in corpus if different, we can find that when add 
 more tag to corpus, the optimal weight is larger.
 2. compare experiment 3 with experiment 6:
  same tag number in corpus, but different jsgf size, we can find that different jsgf size have the 
 same optimal weight.

• need to do

• tag Probability should test add the weight(hanzhenglong) and handover to hanzhenglong (this week)
• make a summary about tag-lm and journal paper(wxx and yuanb)(two weeks).

RNN LM

• rnn

• test wer RNNLM on Chinese data from jietong-data(this week)
• check the rnnlm code about how to Initialize and update learning rate.
• generate the ngram model from rnnlm and test the ppl with different size txt.(this week)

• lstm+rnn

• check the lstm-rnnlm code about how to Initialize and update learning rate.

Word2Vector

W2V based doc classification

• Initial results variable Bayesian GMM obtained. Performance is not as good as the conventional GMM.(hold)
• Non-linear inter-language transform: English-Spanish-Czch: wv model training done, transform model on investigation

Knowledge vector

• Knowledge vector started

• begin to code

Character to wordr

• Character to word conversion(hold)

• prepare the task: word similarity
• prepare the dict.

Translation

• v5.0 demo released

• cut the dict and use new segment-tool

QA

deatil:[1]

Spell mistake

• retrain the ngram model(caoli)

improve fuzzy match

• add Synonyms similarity using MERT-4 method(hold)

improve lucene search

• using MERT-4 method to get good value of multi-feature.like IDF,NER,baidu_weight,keyword etc.(liurong this month)

Multi-Scene Recognition

• handover to duxk(this week)

XiaoI framework

• give a report about xiaoI framework
• new inter will install SEMPRE

patent

• GA-method improve the QA(this week)