“ASR Status Report 2017-8-21”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（5位用户的8个中间修订版本未显示）
第6行： 第6行：
 
|Xiaofei Kang
 
|Xiaofei Kang
 
||  
 
||  
−
*  
+
* Absence
 
||  
 
||  
−
*  
+
* finish experiments on 5 recorded speech.
 +
* Improve and test the human test website.
 +
* Learn 4 papers from lantian : about speaker recongnition and deep speaker feature.
 
|-
 
|-
  
第25行： 第27行：
 
|Yanqing Wang
 
|Yanqing Wang
 
||  
 
||  
−
*  
+
* explore how the pruning method influence the ( distribution of ) output of the network: [http://cslt.riit.tsinghua.edu.cn/mediawiki/images/d/d4/Abs_pos_neg.pdf result]
 +
* after retraining, the distribution may [http://cslt.riit.tsinghua.edu.cn/mediawiki/images/8/88/1_RE.pdf reappear].
 
||
 
||
−
*  
+
* continue on exploration on the network's sparse structure.
 
|-
 
|-
  
第34行： 第37行：
 
|Ying Shi   
 
|Ying Shi   
 
||  
 
||  
−
*  
+
* crawler program [finished]
 +
* tibetan asr system baseline (19.46%)
 
||  
 
||  
−
*  
+
* multilingual decoding
 +
* maybe I can help Zhiyong Zhang to do some work about TTS
 +
* check toolkit code(check data website and codemap) and check it into git
 
|-
 
|-
  
第52行： 第58行：
 
|Lantian Li   
 
|Lantian Li   
 
||  
 
||  
−
*  
+
* Speaker segmentation
 +
** Speaker change points detection + K-means clustering.
 +
** GMM clustering.
 +
** Clean up codes.
 +
* Prepare IS2017 presentation.
 
||
 
||
−
*  
+
* Attend IS2017.
 
|-
 
|-
  
第61行： 第71行：
 
|Zhiyuan Tang  
 
|Zhiyuan Tang  
 
||  
 
||  
−
*  
+
* 1. align the candidate speech (fbank) with phone labels using nnet3-align-compiled (almost finished); 2.analyse the alignment with rhythm, tone, tune, for Parrot system, (revised goodness of pronunciation), to be done.
 +
* collecting material (PPT) for Kaldi toolbook.
 
||
 
||
−
*  
+
* analyse the alignment with rhythm, tone, tune, (revised goodness of pronunciation).
 +
* toolbook writing
 
|-
 
|-
  

2017年8月28日 (一) 14:28的最后版本

Date People Last Week This Week
2017.8.21 Xiaofei Kang
  • Absence
  • finish experiments on 5 recorded speech.
  • Improve and test the human test website.
  • Learn 4 papers from lantian : about speaker recongnition and deep speaker feature.
Miao Zhang
  • Prepare the data and finish experiments on 5 recorded speech.
  • Finish the human test website(include 20 styles), express my apprecation to Shuai sister!
Yanqing Wang
  • explore how the pruning method influence the ( distribution of ) output of the network: result
  • after retraining, the distribution may reappear.
  • continue on exploration on the network's sparse structure.
Ying Shi
  • crawler program [finished]
  • tibetan asr system baseline (19.46%)
  • multilingual decoding
  • maybe I can help Zhiyong Zhang to do some work about TTS
  • check toolkit code(check data website and codemap) and check it into git
Yixiang Chen
Lantian Li
  • Speaker segmentation
    • Speaker change points detection + K-means clustering.
    • GMM clustering.
    • Clean up codes.
  • Prepare IS2017 presentation.
  • Attend IS2017.
Zhiyuan Tang
  • 1. align the candidate speech (fbank) with phone labels using nnet3-align-compiled (almost finished); 2.analyse the alignment with rhythm, tone, tune, for Parrot system, (revised goodness of pronunciation), to be done.
  • collecting material (PPT) for Kaldi toolbook.
  • analyse the alignment with rhythm, tone, tune, (revised goodness of pronunciation).
  • toolbook writing



Date People Last Week This Week
2017.8.14 Xiaofei Kang
  • Recording 35 people audio, located in /work7/zhangmiao/speaker/wavdata/data_new
  • Learn the new test website from zhangmiao
  • Go home with my mom, and come back on Friday night.
Miao Zhang
  • Recording work
  • Test website's data preparation
  • check the linear chapter
  • Continue to record
  • do experiments on recorded speech if possible
  • check the NN chapter
Yanqing Wang
  • TRP uploaded.
  • explore the importance of sparseness structure:
    • After pruning, initialize non-zero values randomly, train.
    • train nnet with 177-dimension hidden layer.
    • result
  • continue exploring the values of trained nnet.
Ying Shi
  • general codeMap finished(kazak)
  • crawler program delayed(Most of the kazakh website is down. I will cralw data from overseas websites)
  • collect more Unicode. such as Tibetan, Mongolia.
  • crawler kazak data from overseas websites.
Yixiang Chen
  • Study English and help Lantian do some Exps.
Lantian Li
  • Visualization and quantification for d-vector [1].
    • phone-aware and phone-blind.
    • within speaker variation and between speaker variation.
  • Speaker segmentation Exps.
  • Finish speaker segmentation Exp.
  • Prepare IS17 presentation.
Zhiyuan Tang
  • reorganize auto-scoring system, next ???
  • collecting material (PPT) for Kaldi toolbook.
  • prefer to rewrite the scoring part.
  • toolbook writing
取自“http://index.cslt.org/mediawiki/index.php?title=ASR_Status_Report_2017-8-21&oldid=28469”