PARAFAC Analysis of 3-D Tongue Shape

Post on 31-Dec-2015

47 views 0 download

Tags:

description

PARAFAC Analysis of 3-D Tongue Shape. Yanli Zheng, Mark Hasegawa-Johnson ECE Department University of Illinois at Urbana-Champaign. Part I. Background. WHY is the factor analysis of tongue shape meaningful ? Speech Motor System. Anatomical View Basic Vowel Diagram. - PowerPoint PPT Presentation

Transcript of PARAFAC Analysis of 3-D Tongue Shape

1

PARAFAC Analysis of 3-D Tongue Shape

Yanli Zheng, Mark Hasegawa-Johnson

ECE Department

University of Illinois at Urbana-Champaign

2

Part I. Background

WHY is the factor analysis of tongue shape meaningful?

1. Speech Motor System

3

Part I. Background (cont. Why?)2. Representing Vowel

• Anatomical View Basic Vowel Diagram

• Frequency Domain

4

Part I. Background

X-ray images

HOW to analysis the vowels in the context of anatomy? 2-D PARAFAC analysis by Richard Harshman(1977)

Measuring Scheme

5

Background (cont. Results of Harshman)

Grids Factors Vowels Loading

Results:Two Factors account for 92% variance.

6

Part I. Background Why is 3D Different from 2D?

1. Linear Source-Filter Theory:• Vowel Quality is Determined by Areas• Area Correlated w/Midsagittal Width

2. Distinguish important in Speech Synthesis

3. Clinic Application

7

Part II. Algorithms Introduction

1. PARAFAC (Parallel Factor Analysis) xijk: tongue shape measurement for ith data point, jth vowel and kth speaker.

aif: fth factor contribution to ith data point

bjf: loading of phoneme j on fth factor

ckf: loading of speaker k on fth facotor

8

Part II. Algorithms Introduction

2. Tucker3 Model(used in the validation of PARAFAC model)

9

Part III. 3-D Factor Analysis of MRI-Derived Tongue Shapes

1. Subjects: 5 subjects successfully imaged (three male speaker: m1,m2, m3; and two female speaker: f1,f2).

2. MRI Image Collection

•T1-weighted•GE Signa 1.5T•3mm slices•24 cm FOV•256 x 256 pixels•Coronal, Axial •11-18 Sounds per Subject.•Breath-hold in vowel position for 25 seconds

10

Part III. 3-D Factor Analysis of MRI-Derived Tongue Shapes 3. Image Viewing and Segmentation: the CTMRedit GUI and toolbox

• Display series of CT or MR image slices

• Segment ROI manually or automatically

• Interpolate and reconstruct ROI in 3D space

11

Part III. 3-D Factor Analysis of MRI-Derived Tongue Shapes 4. PARAFAC Analysis1) 3D-Tongue Shape 2) How to define the

measuring grid?

-40 -30 -20 -10 0 10 2050

60

70

80

90

100

110

120

y-ax

is(fr

om p

oste

rior t

o an

terio

r)

X-Y Plane

x-axis (from right to left)

12

Part III. 3-D Factor Analysis of MRI-Derived Tongue Shapes 4. PARAFAC Analysis(cont.)

3) Result: 2 Factors are extracted, with 83.8729 % variance explained

0

5

-100

10-20

0

20

compenent1

1 2 3 4 5-15

-10

-5

0

5

10coronal view

-10 -5 0 5 10-15

-10

-5

0

5

10sagittal view

0

5

-20

2-10

0

10

compenent2

1 2 3 4 5-10

-5

0

5

10coronal view

-0.5 0 0.5 1 1.5-10

-5

0

5

10sagittal view

13

0.3 0.31 0.32 0.33 0.34 0.35 0.36 0.370.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0.55

0.6Vowel Loadings

hod

had

head

hayed

hid

heedhoed

put shoe

Factor 1

Fact

or 2

0.38 0.39 0.4 0.41 0.42 0.43 0.44 0.45 0.46 0.47-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8Speaker Loadings

f1

f2

m1

m2

m3

Factor 1

Fac

tor

2

14

Part III. 3-D Factor Analysis of MRI-Derived Tongue Shapes 4)Validation of the Result

a) Split-half test (example for f1,f2 and m3)

Correlation Coefficients

Grid Contribution 0.9646

Vowel Loading 0.9279

0.28 0.3 0.32 0.34 0.36 0.38 0.4-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6Vowel Loadings

hod

had

head

hayed

hid

heed

hoed

put

shoe

Factor 1

Facto

r 2

0

5

-10010-20

-10

0

10

compenent1

1 2 3 4 5-15

-10

-5

0

5

10coronal view

-10 -5 0 5 10-15

-10

-5

0

5

10sagittal view

0

5

-1012-10

0

10

compenent2

1 2 3 4 5-10

-5

0

5

10coronal view

-0.5 0 0.5 1 1.5-10

-5

0

5

10sagittal view

15

4)Validation of the Result

b) Check the reliability of the solution• Try different start points, check whether all the solutions

converge to the same solution.

c) Core Consistency Testing (by Rasmus Bro,1998)

1 2 3 4 5 6 7 8-0.2

0

0.2

0.4

0.6

0.8

1

1.2Core consistency 99.9955% (yellow target)

Core elements (green should be zero/red non-zero)

Cor

e S

ize

0 5 10 15 20 25 30-2

-1.5

-1

-0.5

0

0.5

1

1.5

2Core consistency -4.4798% (yellow target)

Core elements (green should be zero/red non-zero)

Cor

e S

ize

16

Degenerated result for 3-factor PARAFAC Model

Correlation Coefficients

1&2 2&3 1&3

Grid Contribution -0.5362 0.9632 -0.6045

0

5

-20-1001020

-20

-10

0

10

20 compenent1

1 2 3 4 5-20

-10

0

10

20coronal view

-10 -5 0 5 10-15

-10

-5

0

5

10

sagittal view

0

5

-20

2-10

0

10

compenent2

1 2 3 4 5-10

-5

0

5

10coronal view

-0.5 0 0.5 1 1.5-10

-5

0

5

10sagittal view

0

5

-10

1-10

0

10

compenent3

1 2 3 4 5-10

-5

0

5coronal view

-0.5 0 0.5 1-10

-5

0

5sagittal view

17

Part IV. Conclusion

• 3-D PARAFAC Analysis of Tongue Shape suggests the “Hierarchical Control”

• This research and the follow-up expected research in the MR Microscopy, and Dynamic Imaging aim to : 1. Provide new anatomical information to speech scientists and

speech pathologists

2. Lay the foundation for future research with disordered populations.