Application of HTK to build continuous speech recognition system (summary)

A HMM-based continuous speech recognition and HTK Toolkit Introduction to the classification of speech recognition system

Person identified: Specific person non-specific

Vocabulary: Large vocabulary in small vocabulary

The way of speaking: solitary word connecting word continuous word

Language: Chinese, English, French ...

We do the non-specific people's vocabulary continuous Chinese speech recognition system

Also called Chinese voice dictation machine

Hidden Markov models (Hidden Markov model)

The Hidden Markov model is a kind of Markov chain, its state cannot be observed directly, but it can be observed by observing vector sequence that each observation vector is represented by some probability density distribution, and each observation vector is produced by a state sequence with response probability density distribution. Therefore, the hidden Markov model is a double stochastic process----a hidden Markov chain with a certain number of States and a set of display random functions.

~o <VecSize> <MFCC_0_D_A>

~h "Proto"

<BeginHMM>

<NumStates> 5

<State> 2

<Mean> 39

0.0 ... 0.0

<Variance> 39

1.0 ... 1.0

<State> 3

<Mean> 39

0.0 ... 0.0

<Variance> 39

1.0 ... 1.0

......

<TransP> 5

0.0 1.0 0.0) 0.0 0.0

0.0 0.6 0.4) 0.0 0.0

0.0 0.0 0.6) 0.4 0.0

0.0 0.0 0.0) 0.7 0.3

0.0 0.0 0.0) 0.0 0.0

<EndHMM>

A HMM model

The HTK Toolkit includes:

Data Preparation Tools

Hdman, Hcopy, hled, Hsgen, Hbuild, Hlstats, Hparse

Model Training and optimization tools

Herest, Hinit, Hrest, hhed, HCOMPV

Identification tool

Hvite

Performance evaluation Tools

HResults, HRec

Two-building continuous speech recognition system

Data preparation

Definition syntax

$word =a|ai|an|ang|ao|ba|bai|ban|bang|

... | Silence;

(Sent-start < $word > Sent-end)

Building an acoustic model

B

P

M

F

D

T

N

L

X

Zh

Ch

Sh

Z

C

...

Last updated to context-sensitive acoustics model

Z+uo

Z-uo

H+ao

H-ao

N+a

N-a

Sh+i

Sh-i

L+i

Sh-ang

Y+ou

Y-ou

D+e

...

Corpus

Sentence 010001: Work number One

Sentence Spell:zuo4 pin3 yi1 Hao4

Sentence 010002: That's a tree for the race.

Sentence spell:na4 shi4 li4 zheng1 shang4 you2 de0 yi1 zhong3 shu4

Sentence 010003: Straight dry

Sentence Spell:bi3 Zhi2 de0 gan4

Sentence 010004: straight branches

Sentence Spell:bi3 Zhi2 de0 zhi1

Sentence 010005: What's it doing?

Sentence spell:ta1 de0 gan4 ne0

Sentence 010006: Usually the height of the Zhang

Sentence spell:tong1 chang2 shi4 zhang4 ba3 gao1

Sentence 010007: As if to be artificial

Sentence Spell:xiang4 shi4 jia1 yi3 ren2 gong1 shi4 de0

Sentence 010008: within one Zhang

Sentence spell:yi1 Zhang4 yi3 nei4

Sentence 010009: Absolutely no side branches

Sentence spell:jve2 Wu2 pang2 zhi1

Sentence 010010: What about all its sluggishly?

Sentence spell:ta1 Suo3 you3 de0 ya1 zhi1 ne0

Sentence 010011: all upward

Sentence spell:yi1 lv4 Xiang4 Shang4

Sentence 010012: and tightly aligned

Sentence spell:er2 qie2 jin3 jin3 Kao4 long3

Sentence 010013: It seems to be artificial.

Sentence spell:ye3 Xiang4 shi4 jia1 yi3 ren2 gong1 shi4 de0

Own recording + online Exchange corpus a total of about 3G corpus

Data feature Extraction

MFCC using hcopy Tools

Data training

Establish a hidden horse model for each primitive

+

Context-Independent training

+

Context-sensitive training

+

Increase the mixing level training

Recognition rate

------------Overall Results------

WORD:%corr=85.71, acc=79.15

==================================