Sunteți pe pagina 1din 9

Automated Earphones

1
Abstract:
Earphones is one of the widely used gadgets these days. It is being used by all
age groups and people are using them everywhere. But there are equal disadvantages
in using them in wrong places and also, listening to high volumes puts one’s health
at great risk. So, the main problem doesn’t lie within earphones but with the volume
at which one is hearing.

The main goal of our project is “To adjust the volume of earphones based on
Ambient Sounds”. For instance, if a person is moving in traffic, his volume has to
be minimum, when there are vehicle honks or ambulance sirens behind, because he
must listen those, else it might lead to an accident. Consider another instance where
a person is sitting at home or walking on road with earphones and construction work
is going on beside, then the volume has to be maximum so that one can listen to the
audio.

For this to happen, we should keep track of ambient sounds. i.e., we need a
microphone to record these external sounds. A microphone will be present along
with earphones which serves the purpose. We need to analyze this recorded sound
and must be able to decide whether the volume of earphones has to be reduced,
increased or must remain same. So, we have to design a network which is able to
make the appropriate decision. This sort of intelligence can be given to our network
by training it with the predefined dataset. This dataset includes various recorded
sounds and the action our network has to take for that particular sound. Once the
network is trained, it is able to categorize the sounds and take appropriate action.
The more we train the network, the high will be accuracy of the decision. Thus the
efficiency of our Automated earphones depends on dataset.

Basically what our network does is, it extracts the some of the audio features
from input audio signal and generates the weights for each of those extracted features
which adds up to a value based on which, it’ll decide the class in which the recorded
signal falls. Once the class is obtained, taking the appropriate action is quite easy.

2
This project once implemented can be widely used, because it is implemented
with low cost and handy to common man. It also reduces lot of accidents which are
being caused due to phone usage in traffic.
Work Done:
 Various audio files(Data set) were collected from different sources for
different classes.
 Audio features were extracted from each audio input using Matlab
functions.
 Neural network was trained using Matlab functions by giving predefined
data set as input.
 Neural Network was designed with the help of Matlab tool which was able
to classify recorded audio files into one of the three output classes.
Progress:
 Training of Neural Network:

As AI based approach must have a preprocessing step of training the


Neural Network with existing data so that any new recording can be fed to
this network which will classify it into one of the classes. This training is
usually done using some technique or algorithm. We started training our
neural network using Back Propagation Algorithm despite using inbuilt
library functions.

 Back propagation Algorithm:


As the name suggests, this algorithm find error values of nodes of a
particular layer and uses this information to calculate the error values
of the previous layer nodes. After optimal number of iterations, errors
in nodes is minimized and accuracy of classification can be increased.
Steps:
1. Initialization of random values to the network parameters (weights).
2. Perform following three steps on the available data set.
3. Compute node values by forward propagation.
4. Error of the last layer is calculated by subtracting the true value
from the obtained value.

3
5. Errors of nodes in the hidden layers is computed by multiplying the
error values of the next layer with the parameters of the current layer
followed by multiplication with the derivative of the activation
function.
6. While performing these steps, accumulate the values obtained by
multiplying node values by next layer error value.
7. Add regularization term to the non bias unit to prevent overfitting.
8. Optimization techniques like Gradient Descent will use these partial
derivative values and cost function values to find the optimal theta
values.

 Training of our neural network when sounds corresponding to two output


classes are present in environment.

 Data set is modified in such a way that input audio files contain
mixture of sounds (corresponding to two different classes) with
varying intensities and assignment of class is done accordingly.
 While testing, when there is more than one sound present in
environment, Classification will be done in favor to that sound which
has high intensity.

 Bringing up GUI for easy user interface(In PC).

4
RESULTS:
- Training neural network

- For vehicle honks

5
- For crowd noises

- For silence

6
-When intensity of crowd noises is more than that of honks

-When intensity of honks is more than that of crowd noises

7
GUI:

Work to be done:
 To bring up a mobile application whose main functionality is to classify
sounds and adjust volume.
 To set up a remote server which receives the recorded audio from mobile
as input and sends the corresponding class as output to phone.
 To complete training of Neural Network without using inbuilt functions.
 To bring up a solution in order to overcome the directionality problem of
microphone.

Comments from previous Presentations:

 To train Neural network without using inbuilt functions.


 To identify importance of extracted audio features.
8
9

S-ar putea să vă placă și