Estimating Multiple Physical Parameters Using Speech Data Shareef Babu
1 Kalluri ,
Ashwin
1 Vijayakumar ,
Deepu
1 Vijayasenan ,
Rita
2 Singh
1
Department of E & C Engg., National Institute of Technology Karnataka, Surathkal, Mangalore, India 2 Language Technologies Institute, School of Computer Science, Carnegie Mellon University, Pittsburgh
[email protected],
[email protected],
[email protected],
[email protected]
Motivation
Effect of Spoken Language
• Speech data contains information about the textual message as well as speaker characteristics
RMSE English Vs Native Language
10
• Can predict characteristics for a person, Linguistic information, Social information, Geographical information
9
English Native Lang
8
• Has an impact on forensic analysis
7
– Eg. Predicting physical traits, geographical information of an anonymous call Cm / Kg
6
Data-set Details
5 4 3
• Collected data-set of 207 Indian speakers
2
• Each speaker contributed speech for 2 minutes
1
• Training – 137 Speakers ( 104 Male + 33 Female )
0
• Testing – 70 Speakers ( 57 Male + 13 Female )
1 Height
• Contains 4 Physical traits i.e., Height, Shoulder size, Waist size, Weight
2 Shoulder
3 Waist
4 Weight
Figure 3: RMSE of English Vs Native language spoke.
Statistics of data-set Physical Trait
Minimum
Height (cm) Shoulder width (cm) Waist size (cm) Weight (kg)
Maximum
147 30 64 39
Mean
188 53 112 107
Standard deviation
168.0 43.5 84.7 64.5
8.5 3.7 7.8 12.1
– The system is robust to Spoken Language
Effect of delta features • 512-dimensional BoW representation is extracted and input to the SVR
Metric
Block Diagram
RMSE
Training Data
STFT log mag
BoW
SVR Training
Feature
Height Shoulder Waist width size
log |ST F T | log |ST F T | + δ
6.6 6.6
2.7 2.6
7.2 7.1
Weight
9.2 8.9
• By adding delta and double delta features, the performance of the regressor improves • Achieved a reduction in the a-posterior variance by 2cm for height, 1cm for shoulder-width and by 3kg for weight estimation
K -Means Comparing RMSE values for the three sub groups .
STFT log mag
Test Data
BoW
SVR Prediction
Figure 1: Block Diagram for Estimating the Physical Parameters
Experiments and Results
Physical Characteristic
p