Speech Project Week 3

Download Report

Transcript Speech Project Week 3

1
專題研究 (3)
Viterbi Decoding
Triphone Acoustic Model
Prof. Lin-Shan Lee, TA. Yun-Chiao Li
2
Viterbi Decoding
03.04.mono0a.viterbi.sh
04.04.tri1.viterbi.sh
Viterbi Decoding
3



Instead of using WFST, we use Viterbi now
Converted Kaldi Acoustic model to HTK by Vulcan
(02.02.convert.htk.feat.sh)
Convert the acoustic model
from Kaldi to HTK
Viterbi Decoding
4
Using the dev set to find the
best acoustic weight (acwt)
5
Triphone Acoustic Model
04.01~04.04
Triphone Acoustic Model
6

In monophone acoustic model,
 ㄅ、ㄆ、ㄇ

In triphone acoustic model,
 ㄅ-ㄆ-ㄇ

they use their own model
is a model
There will be too many model and lack of
training data
Decision Tree
7

Use decision tree to tie similar models together
04.01.tri1.train.sh (1/3)
8

It is very similar to 03.01
04.01.tri1.train.sh (2/3)
9
04.01.tri1.train.sh (3/3)
10
11
Homework
bash 04.01.tri1.train.sh
bash 04.02.tri1.mkgraph.sh
bash 04.03.tri1.fst.sh
bash 04.04.tri1.viterbi.sh
Some Helpful References
12
“使用加權有限狀態轉換器的基於混合詞與次詞
以文字及語音指令偵測口語詞彙” – 第三章

 https://www.dropbox.com/s/dsaqh6xa9dp3dzw/wfst
_thesis.pdf

Check HDecode, HLRescore in HTK Book