專題研究 (2) HINIT + HREST + HEREST

專題研究 (2)HINIT + HREST + HEREST

Prof. Lin-Shan Lee, TA. Yun-Nung Chen

1

HInit + HRest

Acoustic Model Initialization

2

Feature Extraction - MFCC3

Example of MFCC4

Acoustic Model

Hidden Markov Model/Gaussian Mixture Model Acoustic model unit : Initial Final set 3 states per model 1 Gaussian mixture per state

We will increase Gaussian mixture next time Example

5

5

Example of HMM6

Flow Chart

7

7

Phone Model Training8

Train a phone model “ai” Label data

Example

HInit – find a starting point HRest – climb hill

HERest

Acoustic Model Training9

Acoustic Model Training10

Add “ai” into “hmm.set.no_ai” Other phones have been trained

HERest Also hill climbing Consider whole utterance Do several times

sil liou (六) san (三) #er (二) sil

MFCC features

sil

Lable:

sil

liou san #er

HInit + HRest + HERest

Appendix11

HInit12

Hinit -A print command

line -T 1 trace flag to 1 -C ./config/train.cfg set config file -l ai set segment label -o ai hmm def output -S ./scp/train.scp set file of features -L /share/data/TrainingLabel/ set input label

dir proto hmm file

HRest13

HRest -A print command

line -T 1 trace flag to 1 -C ./config/train.cfg set config file -l ai set segment label -S ./scp/train.scp set file of features -L /share/data/TrainingLabel/ set input label dir ai hmm file

HERest14

HERest -A print command

line -T 1 trace flag to 1 -C ./config/train.cfg set config file -S ./scp/train.scp set file of features -L /share/data/TrainingLabel/ set input label dir -H hmmset.mmf load hmm file -M hmm_herest write hmm file

dir ./list/phone.lst hmm list

HInit + HRest + HERest

Homework15

To Do16

Acoustic model initialization HInit +HRest

Train phone model “ai” Acoustic model re-training

HERest HTKBook Chap 8.1 – 8.5

Saving your time…

About Linux17

To Save Your Time …18

Install Cygwin Ref: 工作站使用說明

Install HTK on your PC Download HTK from 課程網站

Tips (cygwin)19

選擇你要的套件 :