ASR:2015-01-19

From cslt Wiki
Jump to: navigation, search

Speech Processing

AM development

Environment

  • May gpu760 of grid-14 be something wrong. To be exchanged.
  • grid-11 often shutdown automatically
  • grid-2/grid-10 have replaced the CPU fan.
  • Add one hard disk to cuda.q machines.

Sparse DNN

RNN AM

Dropout & Maxout & retifier

  • Drop out
  • MaxOut && P-norm(+)
  • Need to solve the too small learning-rate problem
    • Add one normalization layer after the pnorm-layer
    • Add L2-norm upper bound
  • hold

Convolutive network

  • Convolutive network(DAE)
  • Feature extractor
    • Technical report to draft, Yiye Lin, Shi Yin, Menyuan Zhao and Mian Wang

DNN-DAE(Deep Atuo-Encode-DNN)

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

VAD

  • Harmonics and Teager energy features.
  • MPE training
  • Test only Harmonic feature

Speech rate training

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

Speaker ID

Language ID

Voice Conversion

  • Yiye is reading materials
  • HOLD


Text Processing

LM development

Domain specific LM

  • LM2.1
  • mix the sougou2T-lm,kn-discount continue
  • train a large lm using 25w-dict.(hanzhenglong/wxx)
  • find the problem in asr result
  • the model will finish (Tuesday)

tag LM

  • Tag Lm
  • tag Probability should test add the weight(hanzhenglong) and handover to hanzhenglong ("this month")
  • run a tag demo(this week)
  • paper
  • paper submit this week.
  • similar word extension in FST
  • find similarity word using word2vec,word vector is training.
  • set the weight for word
  • set a proper test set
  • write a draft of a paper

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

  • Knowledge vector
  • Make a proper test set.
  • use text information and train word vector together.
  • Modify the object function and training process.
  • try to train on the whole data set
  • result
  • 0.745->0.79, using yago for training.

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

  • review related paper

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 ..
  • extract more features about lexical, syntactic and semantic to improve re-ranking performance.

context framework

  • code for organization
  • change to knowledge graph

query normalization

  • using NER to normalize the word
  • new inter will install SEMPRE