Acoustic Model

Continuous HMM
HTK format
Gender-Dependent (GD) + Gender-Independent (GI)

List of Acoustic Models

#states #mixtures gender

monophone 129 4, 8, 16 GD, GI

triphone 1000 1000 4, 8, 16 GD

triphone 2000 2000 4, 8, 16 GD, GI

triphone 3000 3000 4, 8, 16 GD

PTM triphone 3000/129 64 GD, GI

List of Japanese Phones

a i u e o a: i: u: e: o: N w y

p py t k ky b by d dy g gy ts ch

m my n ny h hy f s sh z j r ry

q sp silB silE (pauses)

Training...ASJ (Acoustical Society of Japan) databases

Phonetically-Balanced sentences (ASJ-PB)
Newspaper Article Sentences (ASJ-JNAS)

20K sentences / 132 speakers for each gender

Acoustic Analysis

A/D 16kHz,16bit

frame shift 10ms

analysis MFCC (12-th order)

LogPow

CMN done for whole utterance

pattern: MFCC + $\Delta$ MFCC + $\Delta$ LogPow (25 variables)

HMM

left-to-right 3 states (excluding initial & final)
decision tree-based clustering:
(logical triphone 21000) $\rightarrow$ (physical triphone 8000)

PTM (Phonetic Tied-Mixture) model

Gaussian distribution set $\leftarrow$ 64-mixture monophone (129 states)
State set $\leftarrow$ 3000-state triphone

Next: Lexicon Up: Specification of Modules Previous: Specification of Modules

Tatsuya Kawahara
5/31/2000

	#states	#mixtures	gender
monophone	129	4, 8, 16	GD, GI
triphone 1000	1000	4, 8, 16	GD
triphone 2000	2000	4, 8, 16	GD, GI
triphone 3000	3000	4, 8, 16	GD
PTM triphone	3000/129	64	GD, GI

A/D	16kHz,16bit
frame shift	10ms
analysis	MFCC (12-th order)
	LogPow
CMN	done for whole utterance