The document presents a final year project on a voice-controlled wheelchair developed by students at Bahir Dar Institute of Technology, detailing components such as voice and speaker recognition systems, control mechanisms through Arduino, and ultrasonic sensors. The project utilizes Mel-Frequency Cepstral Coefficients (MFCC) for feature extraction and artificial neural networks for recognition, achieving high accuracy rates. Challenges faced during the project included limited resources and poor supervision, but the final design is deemed feasible for controlling wheelchair movement.

1. BAHIR DAR INSTITUTE OF TECHNOLOGY
SCHOOL OF ELECTRICAL ENGINEERING
FINAL YEAR PROJECT PRESENTATION :
VOICE CONTROLLED WHEELCHAIR
Members:
Matebie Gashu (500/02)
Mengistu Baye (1307/03)
Mignot Ayana (1373/03)
Mubarek Kebede (14bb/03)
Supervisor :
Mr. : Enyachewu

2. PRESENTATION OUTLINE
 introduction mengistu
 Voice recognition Mubarek
 Speaker recognition
 control system Mignot
 Result and conclusion matebie

3. INTRODUCTION
 Controlling device with speech is used for many
application
 To recognize speech and speaker first the
feature of the speech must be extracted
 the best way to extract speech feature is
MFCC
o After feature extraction we classify the speech
and speaker using artificial neural net work

4. OBJECTIVE OF THE PROJECT
 General Objective
 Design control system of voice controlled
wheelchair.
 Specific Objectives
 To develop speech and speaker recognition
system
 To develop c code for Arduino which control
motor
 To interface Ultrasonic sensor to Arduino

5. .
Literature review
Study mathematical
modeling
Training with neural
network
Do speaker recognition
Voice data collection
Feature extraction
Is test
good?
predict
Do speech recognition

6. Voice recognition

7.  Voice recognition part of our project mainly done
in semester project
 what we did is improving efficiency

8.  Ways of efficacy improvement
 extracting additional feature
 short time energy
 Zero crossing point
 Adjusting good recording time
 And adjusting suppling frequency

9. Speaker recognition

10.  The speaker recognition part is used for change
the system from speaker independent to speaker
dependent

11. CAPTURE AND PREPROCESSING
 Get the audio signal.
 Make suitable for feature extraction
Capture
Silence
Removal
Pre-Emphasis
Framing
Windowing

12. FEATURE EXTRACTION
 Transform the input audio signal into a
sequence of acoustic feature vectors
 MFCC : Mel Filter Cepstral Coefficients
as Feature
 Perceptual approach
 Human Ear processes audio signal in Mel
scale
 Mel scale : linear up to 1KHz and
logarithmic after 1KHz
 MFCC gives distribution of energy in Mel
frequency band
 Calculated for each frame
Fourier Transform
Mel Filter
Log
IFT : DCT
Energy and Deltas

13. COMPOSITION OF FEATURE VECTOR
 20 MFCC Features
 20 Delta MFCC

14. LEARNING PHASE
Artifial neural network (ANN)

15. DESIGN CONSIDERATION OF ANN
Sigmoid activation function
Activation function
Number of input unit
Number of hidden
layer
Number of units per
layer
Number of out put

16. DESIGN CONSIDERATION OF ANN
 The number of input units are the
the size of the external tigers
 in our case the input units are
40 coeffecentes
Activation function
Number of input unit
Number of hidden
layer
Number of units per
layer
Number of out put

17. DESIGN CONSIDERATION
The number of hidden layer affect
learning acurecy
High number hidden layer lead
over fit
High mathematical cost
Small number of hidden layer may
unable to learn
Activation function
Number of input unit
Number of hidden
layer
Number of units per
layer
Number of out put

18. DESIGN CONSIDERATION
 Number of units per layer
 High number of unit per layer
 running cost increase
 It may lead over fit
o Small number of units per layer
 Leads under fit
Activation function
Number of input unit
Number of hidden
layer
Number of units per
layer
Number of out put

19. DESIGN CONSIDERATION
 number of out put unit
the number of the out put unit
in our case 2
default and
User
Activation function
Number of input unit
Number of hidden
layer
Number of units per
layer
Number of out put

20. MODEL
 Based on the above design criteria
We select our model

21. FEEDFORWARD
 feedforward are used to propagate the input
forward to the hypothesis

22. COST
 The root mean square error (RMSE) is
a frequently-used measure of the
differences between values predicted by a
model or an estimator and the values
actually observed from the thing being
modelled or estimated

23. BACK PROPAGATION AND GRADIENT
 Backpropagation is used for propagating the
error from out put to input
 Backpropagation performs a gradient descent
within the solution's vector space towards a
global minimum.
 The gradient from the backpropagation given to
the fimncg bulletin fanction
 Fmincg updating the weights according to the
cost and out put the optimized weights

24. ANN ALGORITHM

25. Hard ware control
design part

26. BLOCK DIAGRAM
MIC
VOICE
RECOGNITION
IC
Arduino
RIGHT
MOTOR
BATTERY
LEFT
MOTOR
H_bridge
Ultrasonic
Alarm and
lighting

27. PARTS OF THE SYSTEM
 In our real system we will use
HM2007 DSP KIT
 It has external interface circuit
 It is dual processing unit
 Good relatively in its speed
 Since can't found in local market
We use normal pc
VOICE RECOGNITION
IC
CONTROL UNIT
OBSTACLE
SENSOR
MOTOR
DIRECTION
CONTROL
MOTOR SPEED
CONTROL

28.  Control unit are the hart of all
control system
In our case we use Arduino uno
 It has 6 analog input and
14 digital i/o
 It has serial communication pin
from external device
VOICE RECOGNITION
IC
CONTROL UNIT
OBSTACLE
SENSOR
MOTOR
DIRECTION
CONTROL
MOTOR SPEED
CONTROL

29.  Obstacle sensor detect presence or
absence of obstacle
In our case we use ultrasonic sensor
VOICE RECOGNITION
IC
CONTROL UNIT
OBSTACLE
SENSOR
MOTOR
DIRECTION
CONTROL
MOTOR SPEED
CONTROL

30.  H-bridge used to control the
direction of the motor
VOICE RECOGNITION
IC
CONTROL UNIT
OBSTACLE
SENSOR
MOTOR
DIRECTION
CONTROL
MOTOR SPEED
CONTROL

31. Current
• One switch is closed in each leg of the "H"
• One switch is open in each leg of the "H"

32. Current
• One switch is closed in each leg of the "H"
• One switch is open in each leg of the "H”
turn in the Other Direction

33.  The speed of the dc motor can
be controlled using PWM from
Arduino
VOICE RECOGNITION
IC
CONTROL UNIT
OBSTACLE
SENSOR
MOTOR
DIRECTION
CONTROL
MOTOR SPEED
CONTROL

34. Start
serial
data
avail
able
Data
range
Is
data<mi
n data?
Read sensor
data
Data={1,2,3,4,5}
Data={6,7,8}
Change
direction
Chang speed
stop
no
yes

35. RESULT AND DESICCATION
 We develop speech and speaker recognition system
 In last we train our data set and gate good accuracy
 Voice recognition 97 for training 88% for test
 Speaker 99 % for training 94 for test
 And we design the control system
 Final we stimulate and we do well

36. PROBLEM FACED
 In this project time we face many problems
 there was no lab which for doing the project so
there was deficiency of computer
 Adviser even if their is there is no follow up
 bad view of the staff member towards as

37.  we will implement the hard ware
properly
It can improved its efficacy by using
vector machine learning
Additional value can be added like
RF transceiver to control remotely

38. CONCLUSION
 From this work we conclude that
 MFCC is fine for speech feature extraction
 Neural network is good for isolated word recognition
In general we conclude that our project is feasible for
control wheelchair movement control.

39. Thanks