This document outlines a project to develop a system that can detect alertness by analyzing speech signals. The objectives are to design and implement the system on a GPU, STM32E development board, and Android platform. The work plan involves literature reviews, algorithm formulation/testing in MATLAB and conversion to C/Java code. Two algorithms are presented - one using generalized eigenvalues for noise removal and MFCC/LPCC features, the other uses GMM/SVM classifiers. The progress made includes functional MATLAB and C/C++ code. Future work involves implementation on GPU, STM32E board and Android app to compare results with other algorithms.

1. Detection of Alertness Based on
Analysis of Speech Signal
Pulak Sarangi
Ojaswa Anand
Induja Sreekant
Bibek Kabi
Under the Guidance of
Prof. Aurobinda Routray
Department of Electrical Engineering
Indian Institute of Technology Kharagpur

2. Objectives
• Design and Develop System Capable of
detecting alertness of a person by analyzing
the speech signal
• Implementation on GPU
• Implement the system on STM32E
development board
• Implementation as app on Android 4.2
(Jelly Bean, target API 17)

3. Work Plan
Week 1
•
Literature Survey
Week 2
•
Formulation of Algorithm
Week 3
•
Algorithm testing on MATLAB
Week 4
•
•
Conversion of MATLAB code to C code
Conversion of MATLAB code to JAVA code
Week 5
•
•
•
Implementation on GPU
Implementation on STM32E
Implementation on Android platform

4. 1. Model As Implemented in MATLAB and C/C++
S(n)
Recording
Formation of
Henkel
Matrix
Classification
of voiced/ silence
parts based on
energy
Noise
Removal
using SVD
Extraction of
Wavelet
Features
De-framing
for Enhanced
Signal
Framing &
Windowing

5. Selection of Wavelet Features
Enhanced
Speech
Segmentation of
speech signal
into
overlapping
samples
6 level
Decomposition
of signal using
Daubechies
wavelet
Computation of
ratio of 62.51000Hz energy to
the total energy
E(i)

6. Classification
E(i) input
<0.3
Comparis
on with
threshold
Silence
Single Segment
with same pre
& post segment
>0.8
Voiced
Series Segment
with same pre
& post
segment
Single or Series
with different
pre & post
segment

7. PROGRESS
• Fully Functional MATLAB & C/C++ code
• Fully Functional Java Code
• Literature survey for implementation of C/JAVA code onto
Embedded/ANDROID platform and GPU respectively.

8. Results
Voiced
Silence
288
311
186
413
151
448
54
545
Speech Signal

9. 2. Model As Implemented in MATLAB and C/C++
S(n)
Recording
Formation of
Henkel
Matrix
Classification
of voiced/ silence
parts based on
Generalized
Eigenvalue
Noise
Removal
using SVD
Feature
Extraction
(MFCCs,
LPCCs)
De-framing
for Enhanced
Signal
Framing &
Windowing

10. Observation
• After feature extraction instead of independent statistical properties
like mean, standard deviation, kurtosis, etc. covariance property was
taken into consideration, making processing much faster.

11. Results
Speech Signal
Distance between
covariance matrices
4.013
6.831

12. PROGRESS
• Fully Functional MATLAB
• Literature survey for implementation of C/C++ code in GPU

13. 3. Model As Implemented in MATLAB and C/C++
S(n)
Recording
Formation of
Henkel
Matrix
Classification
of voiced/ silence
parts by GMM,
SVM classifier
Noise
Removal
using SVD
Feature
extraction
(MFCCs,
LPCCs)
De-framing
for Enhanced
Signal
Framing &
Windowing

14. PROGRESS
• Fully Functional MATLAB code
• Literature survey for implementation of C/C++ code in GPU

15. PLAN FOR FURTHER WORK
•
•
•
•
Implementation on GPU
Implementation on STM32E development board
Implementation as Android App for Android 4.2(API 17, Jelly Bean)
Comparison of Results with other algorithms

16. Thank You

17. Voiced and Unvoiced Sounds
• Fundamental difference :
o Vibrations of the vocal cords produce voiced sounds.
o Rate at which the vocal cords vibrate dictates the pitch of the sound.
o Unvoiced sounds do not rely on the vibration of the vocal cords.
o Unvoiced sounds are created by the constriction of the vocal tract.
o Vocal cords remain open and the constrictions of the vocal tract force air out to produce
the unvoiced sounds
• The fundamental frequency of voiced segments is ranged from 60-500Hz
• The ratio between the energy of the bands between 62.5 Hz and 1000Hz to that of all bands
is computed and used in our algorithm as the fundamental parameter in formulating the
V/UV decision.

18. Literature Review
• Speech Enhancement using Singular Value
Decomposition(SVD)
• Wavelet based Voiced/Unvoiced Classification
Algorithm