Page 42 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

26

There are two common activation functions to use with networks such as the Softmax
function and the rectified linear activation function (ReLU). In this research, we will
choose Softmax as the activation function because it is suitable for classification which
it exploit the relation between input data by put these input follow probabilistic distri-
bution with values that sums to one for optimal output especially for multi-class classi-
fication.
(1)

2.3 Spectrogram

Since the above deep learning model using CNN algorithms only accept imagine as

sound as input by visually them using spectrogram.
The spectrogram is utilized to visualize the length, frequency, and intensity of an
audio signal present in a particular waveform. The spectrogram is computed by a for-
mula called Short-time Fourier transform which is used to determine the sinusoidal fre-
quency and phase content of a signal as it changes over time.
A Spectrogram of a sound will represent it in frequency domain with three parame-
parameters, the x-axis is time and y-axis
are frequency, the magnitude is higher as the bold of the color.

Fig. 4. Sound of a gun display in waveform(above) and spectrogram,
(under) firing in single shot

3 Implementation and Analysis

3.1 Training process

Dataset: -consuming part when training a model.
The research was tasked to collect samples for 4 labels: scream, glass breaking sound,
gunshot and environment. The first label is environment that is the easiest sound we
can collect in real life. That label contains about 100 samples about normal activities
such as talking, laughing, singing and quiet background noise. Another label is scream

CITA 2023 ISBN: 978-604-80-8083-9