Presentation is loading. Please wait.

Presentation is loading. Please wait.

專題進度報告 資工四 B93902054 洪志豪 資工四 B93902076 林宜鴻.

Similar presentations


Presentation on theme: "專題進度報告 資工四 B93902054 洪志豪 資工四 B93902076 林宜鴻."— Presentation transcript:

1 專題進度報告 資工四 B 洪志豪 資工四 B 林宜鴻

2 Flow chart

3 Language Model 1.Normalize(UTF-8) -> normalize_utf8
2.From UTF-8 to Big5 -> iconv 3.parse according to dictory -> textSegment 4.Build language model -> ngram-count

4 Acoustic Model - HERest
1.meaning:Adjust HMM model 2.command: HERest –C herest.cfg –S train.scp –L /share/data/TrainingLabel –H hmmset.mmf rcd.lst 3.Run more can get higher accuracy

5 Combine Acoustic and Language Model(1) pre-requirement
1.change the bigram.lm from Big5 to ascii 2.command: /share/program/big52ascii/chenc bigram_Big5.lm > bigram_ascii.lm 3.replace <s> with !=ENTER </s>with !=EXIT A useful command to replace: ->:1,$ s/<s>/!ENTER/g :1,$ s/<\/s>/!EXIT/g

6 Combine Acoustic and Language Model(2) HBuild
1.meaning:change the SRILM form to HTK network form 2.command: HBuild –n bigram_ascii.lm htkword.dic wd_net 3.htkword.dic is the relation between work and pronunciation EX:[B7C2] k_e e 4.wd_net is that word_net that we need EX:I=9996 W=[A64C][C3D2] J= S=9551 E=2135 l=-7.52

7 Combine Acoustic and Language Model(3) HBuild(Cond.) - Questions
htkword.dic has no word that corresponds to <unk> <sol>: don’t add flag –u when training language model, but is it appropriate?

8 Decode-HVite(1) 1.meaning:decode testing corpus 2.command:
HVite –t 250 –H hmmset.mmf –S testing.scp –C hvite.cfg –w wd_net -i result.mlf –p 0.0 –s 0.0 htkword.dic rcd.lst 3.result is in result.mlf EX: !ENTER [B37D]

9 Decode-HVite(2) - Questions
If we use HERest to train hmmset.mmf with phone.lst,it will error: ->cannot load file n_o <sol>:use rcd.lst instead

10 Divide Mixtures by HHEd
But hmmset.mmf given by TA already has 32 mixtures for each state What config to use? Old DSP hw: the same as what HERest use My trail: it can run without config

11 Initailize Acoustic Model
2 methods for 2 different conditions With labels: use HInit and HRest Without labels: use HCompV Both of them has a similar question

12 Reference 1.HTK_Book.pdf 2.SpeechProject2007Winter.ppt
3.dsp_hw1.ppt(2006_spring_DSP) 4.HTK_Training.pdf 5.SpeechProj board in ptt2


Download ppt "專題進度報告 資工四 B93902054 洪志豪 資工四 B93902076 林宜鴻."

Similar presentations


Ads by Google