Deep Learning based Noise Reduction in FM

Transceivers using the HackRF Platform

Dr V. Kejalakshmi1*, R. Thilagavathy AP(Sr.Gr)2*,
A.R. Avinash3*, N.Dinesh3*, G.Karthik kannan3*
1*
Head of Department,2* ECE Staff,3*UG Scholar
Department of Electronics and Communication Engineering
K.L.N College of Engineering,
Pottapalayam, Sivagangai-630612, Tamil Nadu, India.

Abstract: This paper presents a novel approach to address noise challenges in FM transceivers by incorporating
deep learning algorithms within the SDR platform. Leveraging the power of neural networks, the proposed
system aims to enhance signal quality during transmission by effectively reducing unwanted noise and
interference. Traditional FM transceivers face challenges in maintaining signal integrity in the presence of
noise. This paper introduces a novel solution employing deep learning algorithms within the ADALM-PLUTO
platform to mitigate noise and enhance the overall performance of FM communication systems. In this paper,
deep learning algorithms are applied within the modulation and demodulation processes to effectively reduce
noise in real-time. Also, Convolutional Neural Networks (CNNs) are utilized for spatial feature extraction to
enhance the discrimination between signal and noise components. Recurrent Neural Networks are used to
capture temporal dependencies in the signal, enabling better noise reduction over time-varying channels.
Quantitative and qualitative analyses of the system performance such as Signal-to-Noise Ratio (SNR) and Bit
Error Rate (BER) are computed and the performance metrics are compared with the traditional noise
reduction methods.

Keywords – ADALM-PLUTO, FM transceiver, CNN, Deep learning, RNN, software Defined radio, BER, SNR

I INTRODUCTION
The device known as 'radio' operates within the radio frequency range of the electromagnetic spectrum, capable of
both transmitting and receiving signals. This functionality is crucial for information transfer in various communication
devices, ranging from cellular mobile phones and televisions to automated car door opening systems and transport
vehicles. Despite its importance, traditional hardware-based radio systems face significant limitations, with a notable
lack of cross-functionality and the necessity for physical intervention when configuring the systems. Using such
conventional radio communication systems results in high production costs and limited flexibility in supporting
multiple ranges of radio waveforms.
A device known as an FM radio transceiver functions by transmitting and receiving modulation signals through the air
medium. In the transmitter section, the FM Radio Frequency emits a modulation signal into the air, which is then
captured by the FM Radio Frequency at the receiver. The received modulation signal at the FM Radio Frequency is
subsequently directed to the demodulator input for further processing, ultimately transforming the modulated signal
into an information signal. The researcher's objective is to develop a straightforward FM transmitter using the GNU
Radio Companion, which is linked to the RTL-SDR. GNU Radio, a freely available software, facilitates the creation
of a Software Defined Radio (SDR) without necessitating additional hardware. Implementation of GNU Radio
involves the use of either the C++ or Python programming language.
The Signal to Noise Ratio (SNR) is a standard measure of signal quality for communication systems, defined as the
ratio between the desired signal power and noise power. In the realm of communication, an information signal serves
as a medium susceptible to interference by noise, potentially causing damage to the information signal. Signals
undergoing such interference exhibit a reduction in quality, a characteristic that can be gauged by evaluating the
Signal to Noise Ratio (SNR), expressed in decibels (dB).
In is paper FM transmitter is implemented in SDR platform with adaptive noise cancellation using Deep learning
algorithm. In section II the system model is described. The GNU Radio transmitter block and ADALM-PLUTO SDR
is explained in section III. In section IV deep learning ANC is described. In section V results are displayed. Section
VI conclusion the paper.

1
II SYSTEM MODEL
In FM transmission, a microphone plays a critical role as it captures sound waves and converts them into electrical
signals. As sound vibrations reach the microphone's diaphragm, they cause fluctuations in electrical voltage,
representing the audio signal.

GNU ADALM
Microphone ANC Setup for FM -PLUTO
Transmission SDR

Figure. Block Diagram

Active Noise Cancellation (ANC), is a technology used to reduce or eliminate unwanted ambient sounds by
generating sound waves that are the exact opposite (anti-phase) of the noise. This technology relies on microphones to
detect external sounds, analyzes them, and then generates corresponding anti-noise signals. When these anti-noise
signals are combined with the original noise, they cancel each other out, resulting in a quieter environment for the
listener.
In GNU Radio, a freely available software, facilitates the creation of Software Defined Radios (SDRs) without the
necessity of dedicated hardware. Its implementation relies on either the C++ or Python programming languages. The
merit of GNU Radio lies in its open-source nature, making it a cost-effective solution for various operations. The
GNU software like GNU Radio for FM transmission involves leveraging its signal processing capabilities to generate,
modulate, and transmit audio signals over the airwaves. GNU Radio provides a platform where users can design
signal processing flowgraphs, incorporating blocks for generating audio signals, modulating them onto carrier waves,
filtering, and amplifying as needed. Through compatible hardware interfaces, such as USRP devices or HackRF,
GNU Radio facilitates the transmission of these modulated signals to antennas for broadcast.

. Figure . GNU Radio

The ADALM-PLUTO SDR (Software Defined Radio) is a versatile and affordable platform developed by Analog
Devices. It provides a compact and flexible solution for experimenting with various radio frequency (RF)
applications, including FM transmission. With its wide frequency range (70 MHz to 6 GHz) and high sampling rate
(up to 61.44 MSPS), the ADALM-PLUTO SDR enables users to receive and transmit signals across a broad
spectrum. By leveraging software such as GNU Radio, users can design custom signal processing flows to modulate
audio signals onto carrier waves, filter, and amplify them for transmission via the ADALM-PLUTO's RF output port.
Additionally, the ADALM-PLUTO SDR offers a user-friendly interface and supports various programming
environments, making it accessible for both beginners and experienced radio enthusiasts to explore FM transmission
and other SDR applications.

2
III SOFTWARE DESIGN TRANSMITTER

The software used is GNU Radio with several blocks including Wav File Source, Audio Source, Embedded Python
Block, Multiply Const, Add, Transmit WBFM, Rational Resampler, WX GUI FFT Sink, Osmocom Sink.

Figure . Block diagram GNU Radio Transmitter

WAV File Source block, which allows to read audio data from a WAV file and stream it into flowgraph. Simply drag
and drop the block onto your canvas, specify the file path and any necessary parameters like sample rate or number of
channels, then connect its output to your flowgraph, Audio Source" block. This block enables to capture audio from
computer's audio input devices, such as a microphone or line input, The Multiply Const block in GNU Radio is a
vital tool for adjusting the amplitude of an input signal by a fixed factor.

By applying a constant multiplier to each sample of the input signal, this block facilitates amplification or attenuation,
depending on the value provided. Whether it's boosting signal strength for improved clarity or attenuating to prevent
distortion.To transmit wideband FM (WBFM) signals in GNU Radio, create or import an audio source, modulate it
onto a carrier frequency using an FM modulator, connect to SDR hardware, and run the flowgraph to start
transmission.

The "Embedded Python Block" in GNU Radio allows to write custom Python code directly within flowgraph. This
block is useful for implementing specific signal processing tasks or integrating external libraries seamlessly in GNU
Radio project.The Rational Resampler, employed for signal filtering before transmission, regulating interpolation and
dissemination.

The WX GUI FFT Sink in GNU Radio is a tool for visualizing the frequency spectrum of a signal in real-time. It
displays a graphical representation of the signal's frequency content, helping you analyze its spectral characteristics
easily. The "PlutoSDR Sink" in GNU Radio enables to transmit signals from flowgraph directly to an Analog
Devices PlutoSDR device for wireless communication.

IV SOFTWARE DESIGN ANC

The proposed approach of real-time noise cancellation in FM transmission entails multiple steps. Data preparation
comes first. To create an RNN-based noise suppression model, the researchers want to use a dataset they downloaded
from the University of Edinburgh's Data Share website. This collection of audio files includes recordings made in a
variety of settings, including offices, cafes, and schools, all of which have various kinds of background noise. By
utilising spectral gating, which divides the audio into active and inactive parts, the researchers will pre-process the
raw audio data. This will enable them to concentrate on the pertinent speech signals. Following spectral gating, they
will extract aspects of the spectral contrast from the voice signals. The spectral characteristics of voice signals can be
captured via spectral contrast features, which can also be utilised to detect and suppress noise. The spectral
characteristics of voice signals can be captured via spectral contrast features, which can also be utilised to detect and
suppress noise. The proposed solution has a number of benefits over conventional noise suppression methods. It is
suitable for usage in a variety of situations due to its ability to tolerate a wide range of noise types and settings.
3
The goal of noise reduction is to take out undesirable noise from an audio source. Noise can have a severe effect on
the audio's quality, making it challenging to listen to and comprehend. In the process of spectral gating, an audio signal
is divided into manageable chunks, its frequency content is examined, and the noise-containing frequencies are then
selectively attenuated. A threshold is used by the spectral gating method to distinguish between the desired signal and
the undesirable noise. The amount of noise in the signal is used to calculate the threshold. The goal of the spectral
gating algorithm is to preserve the intended signal while attenuating just the frequencies that are above the noise level.
This method minimises the impact on audio quality while allowing for excellent noise suppression. Spectral gating is a
common method for noise reduction in audio processing that is used in a variety of applications, such as sound
engineering, speech processing, and music production. Windowing, short- time Fourier transform (STFT),
thresholding, and spectral attenuation are the four main components of spectral gating.

Windowing is the division of an audio signal into smaller windows that are then each individually evaluated. The
window size is significant in spectral gating since it affects the analysis's frequency resolution. Greater frequency
resolution is possible with a smaller window size, allowing for more accurate noise frequency detection. The time
resolution is also decreased with a smaller window size, which can make it more challenging to separate the intended
signal from the unwanted noise. The mathematical process known as the short-time Fourier transform (STFT) is used
to convert the audio signal from the time domain to the frequency domain. On each audio signal window, the STFT is
applied. The resulting frequency spectrum gives details on the signal's frequency components' amplitude and phase.
The time resolution is also decreased with a smaller window size, which can make it more challenging to separate the
intended signal from the unwanted noise.

Setting a threshold level is the process of determining the difference between the desired signal and the undesirable
noise. The amount of noise in the signal is often used to calculate the threshold level. The statistical characteristics of
the noise can be used to either manually or automatically set the threshold. The threshold in spectral gating is used to
choose which frequency components should be muted or repressed. A frequency component in a certain segment is
silenced or attenuated if it exceeds the threshold. To reduce the influence on the audio quality while achieving the
appropriate level of noise reduction, the threshold can be changed.
An RNN is a kind of neural network that is capable of processing sequential input, which makes it ideal for handling
time-series data, such audio signals. An RNN's ability to retain a recollection of prior inputs, which enables it to
represent the temporal connections between the input and output, is its fundamental feature. Recurrent connections are
introduced into the network to do this, allowing data to be sent from one time step to the next. By doing this, the
network can keep track of prior inputs and use that knowledge to predict the output more accurately. The recurrent
cell, which accepts an input at each time step and generates an output as well as a hidden state, is the fundamental
component of an RNN.

An RNN (Recurrent Neural Network) can be trained as an autoencoder in Tensorflow for preprocessing tasks. An
autoencoder is a neural network that learns to compress and reconstruct the input data. In the case of an RNN
autoencoder, the network is designed to learn the temporal dependencies in the input sequence, and compress the
sequence into a fixed-length vector called the "latent representation". This latent representation can then be used for
downstream tasks such as classification, prediction, or other processing.

The input sequence is first preprocessed to remove any noise or irrelevant features. The encoder part of the RNN
autoencoder consists of an RNN layer that reads in the input sequence and produces a compressed latent
representation of the sequence. The decoder part of the RNN autoencoder consists of another RNN layer that takes the
latent representation and reconstructs the input sequence. This is done by feeding the latent representation into the
decoder RNN layer and iteratively generating the next output element until the entire sequence is reconstructed.

4
The loss function is used to train the autoencoder to reconstruct the input sequence accurately. The loss function can
be based on the difference between the original input sequence and the reconstructed output sequence, using metrics
such as mean squared error or binary cross-entropy. The autoencoder is optimized using gradient descent or other
optimization techniques to minimize the loss function. This involves updating the weights and biases of the RNN
layers in the encoder and decoder. Once the RNN autoencoder is trained, the latent representation can be used as a
preprocessed input for downstream tasks such as classification, prediction, or other processing. By learning the
temporal dependencies in the input sequence, the RNN autoencoder can effectively preprocess the data and extract the
relevant information, improving the performance of subsequent tasks.

V RESULT AND DISCUSSION

INPUT AUDIO WAVEFORM

TRANSMITTED SIGNAL

VI CONCLUSION

A software-defined radio (SDR) is commonly defined as a radio communication system that utilizes specific software-based
modules for signal transmission instead of relying on hardware-based systems. Deep learning has proven remarkably effective in
real-time noise suppression on the transmission side. Through the implementation of an autoencoder using TensorFlow and
training an RNN model, we achieved a significant improvement in the signal-to-noise ratio (SNR) compared to conventional
techniques like LMS.

This academic research aimed to explore the utilities of a software-defined radio system in modern communication systems and
gain hands-on experience in the practical implementation of such systems. Notably, the research focused on implementing a
software-defined radio communication system on ADALM-PLUTO SDR, a well-known embedded hardware system specifically
developed for software-based FM transmission.

5
VII REFERENCE

[1]. N. F. A. Hakim and I. Kustiawan, "Experimental Study of FM Complex Differentiation Using HackRF," 2021
International Research Symposium On Advanced Engineering And Vocational Education (IRSAEVE), Malang,
Indonesia, 2021, pp. 10-13
[2]. I. Martoyo, P. Setiasabda, H. Y. Kanalebe, H. P. Uranus and M. Pardede, "Software Defined Radio for
Education: Spectrum Analyzer, FM Receiver/Transmitter and GSM Sniffer with HackRF One," 2018 2nd Borneo
International Conference on Applied Mathematics and Engineering (BICAME), Balikpapan, Indonesia, 2018, pp.
188-192
[3]. Design Indoor FM Communication Based on SDR and GNU Radio Using Validated Spectrum Analyzer To cite
this article: B B Harianto et al 2021
[4]. Tan, K., & Wang, D. L. (2019a). Complex spectral mapping with a convolutional recurrent network for monaural
speech enhancement. In 2019 IEEE international conference on acoustics, speech and signal processing(pp. 6865–
6869).
[5]. Tan, K., & Wang, D. L. (2019b). Learning complex spectral mapping with gated convolutional recurrent networks
for monaural speech enhancement.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28, 380–39
[6]. Wang, D. L., & Chen, J. (2018). Supervised speech separation based on deep learning: An overview. IEEE/ACM
Transactions on Audio, Speech, and Language Processing, 26, 1702–1726.
[7]. Gao, F., Wu, L., Zhao, L., Qin, T., Cheng, X., & Liu, T.-Y. (2018). Efficient sequence learning with group
recurrent networks. In Proceedings of the 2018 conference of the north american chapter of the association for
computational linguistics:human language technologies
[8]. A. W., Beerends, J. G., Hollier, M. P., & Hekstra, A. P. (2001). Perceptual evaluation of speech quality (PESQ)-a
new method for speech quality assessment of telephone networks and codecs. In 2001 IEEE international conference
on acoustics, speech, and signal processing. proceedings (Cat. No.01CH37221) (pp. 749–752).