Speech Project
Week-4 Progress Report
B02901054 方為
B02901085 徐瑞陽
Training Flow
Feature Extraction
Monophone
Equal align
即日起 簡章 備索 in lexicon.txt
Gmm align
即，起 的 '一' 拉很長 ， '日' 是短音
tree
produced by gmm-init-mono?
Triphone
Convert-ali
decision tree <-> model
output is tree and model
same as gmm-init-mono
gmm-mixup
VS
gmm-est mixup
Meaning?
WFST
Viterbi
Expriment Report
| # of Iters | # of gaussians | accuracy (%) |
|---|---|---|
| 40 | 500 | 54.96 |
| 40 | 1000 | 54.64 |
| 40 | 1500 | 55.79 |
| 40 | 2000 | 55.1 |
| 40 | 2500 | 53.98 |
| 40 | 3000 | 51.22 |
| 40 | 4000 | 55.9 |
Viterbi Triphone
change # of gaussians
| # of Iters | # of gaussians | accuracy (%) |
|---|---|---|
| 40 | 1000 | 54.64 |
| 70 | 1000 | 55.90 |
| 100 | 1000 | 55.36 |
Viterbi Triphone
change # of iters
Viterbi Monophone
| # of iters | # of gauss | accuracy(%) |
|---|---|---|
| 40 | 1000 | 47.52 |
| 70 | 1000 | 45.59 |
FST Monophone
| # of iters | # of gauss | accuracy(%) |
|---|---|---|
| 40 | 1000 | 34.32 |
FST Triphone
| # of iters | # of gauss | accuracy(%) |
|---|---|---|
| 40 | 1000 | 31.86 |
8 hours ...= =" ,
monophone is 2 hours,
but still lower than monophone!?
Problems
we encountered
Gaussian has
too little data...OAO
The only difference
scp file for reading in feature extraction
Is it ok?
Sth weird happened
in new workstation...
