ASR:2015-02-02

Speech Processing

AM development

Environment

• May gpu760 of grid-14 has been repairing.
• grid-11 often shutdown automatically, too slow computation speed.

RNN AM

• details at http://liuc.cslt.org/pages/rnnam.html

Dropout & Maxout & rectifier

• Need to solve the too small learning-rate problem
• 20h small scale sparse dnn with rectifier. --Chao liu
• 20h small scale sparse dnn with Maxout/rectifier based on weight-magnitude-pruning. --Mengyuan Zhao

Convolutive network

• Convolutive network(DAE)

• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=311
• Technical report to draft, Mian Wang, Yiye Lin, Shi Yin, Mengyuan Zhao

DNN-DAE(Deep Auto-Encode-DNN)

• Technical report to draft, Xiangyu Zeng, Shi Yin, Mengyuan Zhao and Zhiyong Zhang,
• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=318

RNN-DAE(Deep based Auto-Encode-RNN)

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

VAD

• DAE
• Technical report --Shi Yin

Speech rate training

• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=268
• Technical report to draft. Shi Yin
• Prepare for ChinaSIP

Confidence

• Reproduce the experiments on fisher dataset.
• Use the fisher DNN model to decode all-wsj dataset
• preparing scoring for puqiang data
• HOLD

Neural network visulization

• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=324

Speaker ID

• DNN-based sid
• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=327

Language ID

• GMM-based language is ready.
• Delivered to Jietong
• Prepare the test-case
• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=328

Voice Conversion

• Yiye is reading materials
• HOLD

Text Processing

LM development

Domain specific LM

• LM2.X

• mix the sougou2T-lm,kn-discount continue
• train a large lm using 25w-dict.(hanzhenglong/wxx)

• v2.0a adjust the weight and smaller weight of transportation is better.(done)
• v2.0b add the v1.0 vocab(this week)
• v2.0c filter the useless word.(next week)
• set the test set for new word (hold)

tag LM

• Tag Lm

• code is given to jietong .

• similar word extension in FST

• write a draft of a paper
• result (need result)

RNN LM

• rnn

• test wer RNNLM on Chinese data from jietong-data
• generate the ngram model from rnnlm and test the ppl with different size txt.

• lstm+rnn

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

Word2Vector

W2V based doc classification

• data prepare.

Knowledge vector

• run the big data

• continue to test, including paragraph vector and relation.
• result: http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=lr&step=view_request&cvssid=326

• prepare the paper.

Character to word

• Character to word conversion(hold)

Translation

• v5.0 demo released

• cut the dict and use new segment-tool

Sparse NN in NLP

• write a technical report(Wednesday) and make a repoert.

QA

improve fuzzy match

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

improve lucene search

• add more feature to improve search.

• POS, NER ,tf ,idf
• result:P@1: 0.68734335-->0.7763158 1097-->1239 (The number of queries is 1596.)P@5: 0.80325814-->0.8383459 1282-->1338 (The number of queries is 1596.)[1]

• extract more features about lexical, syntactic and semantic to improve re-ranking performance.
• using sentence vector, it doesn't work.

context framework

• code for organization

• change to knowledge graph,and learn the D2R tool and JENA

query normalization

• using NER to normalize the word

• new inter will install SEMPRE