“2014-03-14”版本间的差异

2014年3月14日 (五) 02:38的最后版本

Resoruce Building

• Current text resource has been re-arranged and listed

AM development

Sparse DNN

• Optimal Brain Damage(OBD).

1. GA-based block sparsity

Efficient DNN training

1. Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test. Overfitting?

Multi GPU training

• Error encountered

GMM - DNN co-training

• Initial DNN test done

• tri4b - > DNN (org)
• DNN alignmenment -> tri4b
• tri4b alignment -> DNN (re-train)

  model/testcase              |  test_dev93(cv)       |     test_eval92
    --------------------------------------------------------------
    8400-80000(org)           |    7.41               |      4.13
    --------------------------------------------------------------
    re-train (Keep state #)   |    7.20               |      4.24
    --------------------------------------------------------------
    re-train (Free state #)   |    7.29               |      4.31
    --------------------------------------------------------------

Multilanguage training

1. Pure Chinese training reached 4.9%
2. Chinese + English reduced to 7.9%
3. English phone set should discriminate beginning phone and ending phone
4. Should set up multilingual network structure which shares low layers but separate languages at high layers

Noise training

• Train with wsj database by corrupting data with various noise types

• Almost all training conditions are completed
• Interesting results in multi-conditional training (white + cafe) and test on park/station

AMR compression re-training

• WeChat uses AMR compression method, which requires adaptation for our AM
• Test AMR & non-AMR model

model			wav	amr

xent baseline		4.47	
wav_mpe baseline        4.20	36.77

amr_mpe_lr_1e-5		6.27	8.95
amr_mpe_lr_1e-4		7.58	8.68

amr_xEnt_lr_1e-5	6.89	7.99
amr_xEnt_lr_1e-4	6.61	7.28
amr_xEnt_lr_0.08	5.72	6.20

• Prepare to do adaptation on 1700h
• Prepare to do mixing xEnt test

GFbank

• Finished the first round of gfbank training & test
• The same gmm model (mfcc feature) was used to get the alignment
• Traing fbank & gfbank based on the mfcc alignment
• Clean training and noise test

		clean	5dB	10dB	15dB	20dB	25dB
gfbank		4.22	73.03	39.20	16.41	8.36	5.60
gfbank_80	4.36	74.41	42.94	18.13	8.59	5.85
fbank_zmy	3.97	74.78	44.57	18.80	8.54	5.30

• gfbank + fbank 80 dim training/test

Engine optimization

• Investigating LOUDS FST.

Word to Vector

• Improved wordvector with multi sense

• Almost impossible with the toolkit
• Can think of pre-training vectors and then do clusering

• WordVecteor-based keyword extraction

• Prepared 7 category totally 500+ articles
• A problem fixed to retrieve article words

• Wordvector based on classification

LM development

NN LM

• Character-based NNLM (6700 chars, 7gram), 500M data training done.

• boundary-involved char NNLM training done
• Test on rescoring

• Investigate MS RNN LM training

3T Sogou LM

• 3T + tencent LM combination:

• Combine the 3T voc (11w) and the tencent 8w voca
• re-segmentation
• compute PPL with the 3T and tencent LM
• compute the best mixing weights
• the mixing weight is wrong ....
• if we mix the two by equal weight (0.5/0.5), performance is better than the individual

• 3T + QA model combination

QA Matching

• FST-based matching

• Investigating why openfST union does not lead to a determinizable graph
• Test the pattern label

• TF/IDF weight

• code is done, TF/IDF weight can be used right now.

Embedded development

• CLG embedded decoder is almost done. Online compiler is on progress.
• English scoring is under go

Speech QA

• N-best with entity LM was analyzed

• WER vs QA accuracy is done
• The figure shows that WER and QA accuracy is positively related
• Addding song names and singer names improve performance in most cases
• There indeed some exceptions in the figure that (a) higher WER does not reduce QA necessarily (b) adding entity names does not improve QA
• The results on 媒体文件:Music_QA_wer.pdf

• Class LM QA
• Use QA LM as the baseine
• Tag singer names and song names
• build tag LM
• Using graph integration to resolve the tags
• Adjusting in-tag weight
• Smaller weight produces more entity recognition
• Check if the recognized songs/singers are correct/wrong

1, non-merge
    BaseLine:
           qa-singer-song
    songs       41
    singers     23

2, HCLG-merge
    Weight means the multiplier of the sub-graph entry. 
  (1) LM:1e-5
    weight  0.00000001  0.0001   0.001  0.01    1   10
    songs      20        20      21    19       9    4
    singers    13        13      13    13       2    2