International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.

5, September 2024

A MULTIFACETED AI-DRIVEN RADIO NETWORK

George Sturm

Radio Logos, Pogradec, Albania

ABSTRACT
This paper discusses a unique and revolutionary type of radio network which is driven by various AI
techniques. Besides the traditional features of any FM radio network, Radio Logos has incorporated into
its system a variety of automated features like weather forecasts, earthquake announcements, monetary
exchange rates, listener song requester capability, and the monitoring of equipment. In addition, an
advanced real-time defect monitoring system for RF signals has been developed and implemented to
monitor the network 24/7. These methodologies are based upon AI techniques ranging from the simple to
the advanced.

KEYWORDS
AI, machine learning, FM radio network, RF signals, MPX, RTL-SDR, time-series, big data, pattern
recognition, discriminant analysis, analysis of covariance, logistic regression, NLP, real-time monitoring.

1. INTRODUCTION
Radio Logos is the largest regional FM radio network in the European country of Albania. It
currently consists of five transmission sites (with two more in the implementation stage) each
with its own unique FM frequency (MHz). These sites are interconnected by means of either
microwave links or by IP LAN connections. Practically speaking, this setup gives the listener
uninterrupted and simultaneous broadcast over a wide geographical area of Albania.

The first goal of Radio Logos is to bring quality music and edifying programs to its listeners. To
fulfill this goal, a dedicated staff works tirelessly behind-the-scenes doing traditional studio tasks
such as news research, music selection, interviews, recording, editing, and mounting to produce a
listener-enjoyable product. Much has been made of recent so-called cost-saving innovations such
as Futuri AudioAI (formerly RadioGPT) [1], ChatGPT [2], GPT-4 [3], and speech synthesis
technology [4] that can find stories, write scripts, and even host an AI-generated talk show. While
these are clever techniques that, in fact, can be implemented, the output of the algorithms could
very well be based upon the latest social media trends, the content of which might lack truth and,
thus, be in opposition to the moral ethics of the radio station. In addition, AI generated stories and
so-called talk shows lack the personal touch of real local people who are in daily contact with the
regional audience and know their needs. Thus, it is not unusual that the AI-generated speaker
comes across as just that – a fake person. Nevertheless, AI-generated techniques can and should
be integrated into any radio network, methodologies that, in fact, reduce operating expenses, ease
the burden upon the staff, and create listener-sought programs. Some of these programs will be
presented in Section III.

The second goal of Radio Logos is to maintain uninterrupted and quality broadcasting over its
entire regional network. During the propagation process of the FM signal, all sorts of defects can
happen to distort the final transmission - electrical problems, cable defects, weather-related
issues, equipment failures, etc. To maintain a quality FM signal, these defects need to be

DOI:10.5121/ijaia.2024.15505 67
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
identified and repaired quickly. If not, our listening audience will seek other sources for their
audio enjoyment. Therefore, a highly sophisticated AI-driven real-time defect analysis system has
been developed and implemented. This will be discussed in Section IV.

2. HOW DO WE USE THE TERM “AI-GENERATED”?

Before embarking upon a discussion of AI-generated techniques, it is important for the reader to
know what we mean by this expression.

First, there are several methods that are cited in contemporary literature that are, in the author’s
opinion, different words for basically the same thing – newer terms like AI, Machine Learning,
Deep Learning and Big Data along with older terms like Pattern Recognition and Discriminant
Analysis. Although one might argue that each is a different methodology, they are, in essence, all
attempting to solve the same thing – theoretically modeling a real-life problem and then
empirically estimating the solution to the equations and/or algorithm.

Second, to further emphasize the above point, whether new or old, all the above techniques
involve the use of the same basic tools – computers, algorithms, software, databases,
mathematics, and statistics. Therefore, the idea of saying that we have developed an AI-generated
technique is somewhat misleading since there is no intelligence being created. Likewise, the
expression Machine Learning gives the wrong impression that a machine is somehow learning
something. Nothing could be further from the truth. We are simply using computers with
mathematical and statistical tools, perhaps in a new and clever manner, to solve equations,
algorithms and, ultimately, the problem. Sometimes the solutions are “deep”, that is, they are
complex equations involving many parameters with interaction terms.

Thirdly, the term “Big Data” is a very “hot” topic, so much so that it is now a required course in
many CS programs. Unfortunately, like the topics mentioned above, it is quite misleading. It is
often presented in a manner that gives the listener the idea that “Big Data” is an issue that has
only arisen in recent years. It must be understood that what is defined today to be “Big Data,”
will become “Small Data” in a few years. The reality is that, over the years, each generation of
scientists has been confronted with analyzing what they considered, at that time, to be a large
amount of data. As a result, they had to develop new mathematical, computer, and statistical
techniques to solve the problem in question. While it is true that the “Information Age” continues
to make available for analysis massive amounts of data, this does not change the basic thrust of
ongoing scientific research – not the creation of new intelligence, but the use of computers and
mathematical tools to process the data and arrive at solutions. As a final thought, just because one
has a problem involving “Big Data,” this does not mean that the solution must be “big.” On the
contrary, a good solution to a problem might be something much simpler than overly complex
models involving thousands of parameters. Sometimes, simpler is better.

Finally, Natural Language Processing (NLP) techniques are routinely used in solving many
contemporary problems and fall under the umbrella of “computer and mathematical tools.”
Therefore, our use of the term “AI-generated” methodologies simply means that we are using
mathematical and computer tools to solve our problems. Perhaps we could have chosen as a title
“A radio network driven by computer science and mathematical methods.” However, it seemed
much simpler to use the expression “AI” with the understanding that its meaning is as described
in the points above.

68
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024

3. LISTENER-SOUGHT PROGRAMS
As previously mentioned, radio listeners enjoy hearing regular information programs. Of course,
news broadcasts are at the top of the list and are carefully created by our staff. However, there are
other programs that are AI-generated such as weather reports, exchange rates announcements,
and flash-news involving earthquakes.

3.1. Weather Reports

Using traditional methods, this is a high-cost product since it requires a person to be available at
all hours of the day, especially in the early hours of the morning when such information is highly
sought. However, using AI techniques, such reports can easily be produced, broadcast at any time
of the day, and sound like a live report to the listener.

How is this done? The first step is one labor intensive day for the person who will do the weather
reports. In Albanian, the person will read a variety of introductions, city names, temperatures, and
weather conditions, and a selection of closing comments. These are then edited, processed using
audio tools, and made ready for the next step.

Several minutes prior to the scheduled weather broadcast times, a software program (e.g. written
in Python or R) is activated to automatically read weather related webpages, extract (for today
and tomorrow) the high and low temperatures and weather conditions (e.g. snow, rain, sunshine,
etc.). This is done for a variety of cities that fall in our network coverage. Then another software
program will use this information to select the needed audio files from our inventory of pre-
recorded information, combine these files in the correct order to produce a broadcast ready
weather report. It must be added that the program will randomly select an introduction and
closing from the inventory to give the broadcast a non-monotonic sound. In the end, our listeners
think that we have a person at the studio 24/7 giving weather reports, so “live” these reports
sound.

3.2. Exchange Rates

Perhaps in other countries, such a program would generate little interest. However, a large
percentage of Albanians work outside of the country and send money (mainly Euros and USD)
back to their loved ones. Thus, people have an interest in the daily exchange rates for these two
currencies. The method to produce such programs is identical to the procedure for the weather
reports except that the pre-recorded files include decimal numbers not just integers. In the end, it
is broadcast twice a day and sounds like a live report.

3.3. Earthquakes

There is nothing more that will generate interest from our listening audience than a good
earthquake. How big was it? Where was it located?

First, we read a website that reports Mediterranean area earthquakes in real-time. Second, we
narrow the list down to those earthquakes that have occurred within the last two hours, greater
than 4.0, and are within a 300 km radius from the studio. The latter step involves the use of a
sophisticated trigonometric equation. Third, using a list of major cities that are within the 300 km
radius, we find the nearest city to the epicenter. Of course, all possible introductions, cities, and
strength must be pre-recorded. Fourth, our software program combines the necessary files to
produce a very live-sounding broadcast. Finally, since earthquakes are unpredictable, this

69
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
newscast falls into the category of “flash-news.” That is, whenever it happens, we must be ready
to put out a broadcast. This means that our software program is constantly running in the
background, monitoring the earthquake webpage, and ready to produce the audio report at a
moment’s notice.

3.4. User Song Requests

To our knowledge, we are the only Albanian radio station that allows listeners to request songs
by means of our website or Android app. The listener requests a specific title or songs by a
particular artist, a list of possibilities is returned from our inventory of approximately 5000 songs,
the selection is made, and a few minutes later the song is broadcast over the entire network.
Truly, our network is listener oriented.

Admittedly, the above AI-generated techniques are simple in that they primarily are using NLP
and audio techniques combined with some clever software programming. However, the next
section describes a more sophisticated methodology that involves a large quantity of data,
mathematics, statistics, and time-series analysis.

4. REAL-TIME DEFECT ANALYSIS SYSTEM

Besides the goal of transmitting quality music and edifying programs, another major goal is to
keep the network running smoothly. It is foolhardy to think that one does not need to invest a
great deal of time and money into this effort. Undetected defects in the FM broadcast signal will
obligate the listeners to turn off their radios or switch to another station.

Radio Logos has multiple antenna sites, each transmitting at a different FM frequency. Of course,
a good method to monitor defects is for the operator to simply listen to the radio. However, with
a regional radio network, this is impossible since usually only one frequency will be heard in the
geographical area of the studio. Traditionally, the method for monitoring a remote site was to ask
someone in that area to be your “controller.” Unfortunately, not even a faithful controller would
be able to listen to the broadcast 24/7. As a result, defects could continue for hours, perhaps days,
without being detected. Therefore, a low-cost AI-generated system was needed that would be
capable of monitoring the entire network 24/7.

To do this, we installed at every transmission site an internet connection and a RTL-SDR

(RealTek Software Defined Radio) dongle attached to a Windows-based computer. The RTL-
SDR dongle is an inexpensive device (less than $50) that can be used as a computer-based radio
scanner for receiving live radio signals up to 1766MHz which means that it can monitor any FM
transmission but also the micro-wave signals of our links that interconnect our transmission sites.
GNU radio [5] also has to be installed. Afterwards, using software like Python or R, the device
can be activated.

Figure 1. RTL-SDR Dongle [found on RTL-SDR.com]

RTL-SDR devices have been used in a variety of applications. For example, environmental
monitoring [6], spectrum monitoring [7,8,9,10], radio coverage maps [11], processing

70
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
Electrocardiogram (ECG) signals [12], and remote monitoring of the content of FM broadcasting
[13,14].

However, in our situation we will use the device to mathematically “listen” to the FM signal at
any time during the day and immediately decide if it was defective or not. However, this required
a database of actual transmissions, data from which we could build our corresponding
mathematical models to solve this problem.

4.1. Data Collection and the Feature Vector

Using the RTL-SDR device, over the course of five months, we recorded over 2300 10-second
FM audio samples. For example, at the FM frequency of 88.6 MHz, the DOS type of command
which is passed to the GNU Radio directory was.

rtl_sdr -f 88600000 -g 10 -s 225001 -n 2250010 Capture.dat

These samples (2349) were recorded from our five FM transmission sites. The table below shows
the number of samples collected at each site, its frequency (in MHz) and site name. Our policy
was to use 80% of our data for training purposes (1879) and 20% for testing (470).

Table 1. Transmission sites and sample sizes

88.6 94.2 96.3 100.3 103.5

Gramsh Pogradec Elbasan Librazhd Cevernake
447 704 605 351 242

It is imperative to realize that each of the above RTL-SDR generated files is an IQ data file of
size 4395KB. From these files, we must extract our multi-dimensional feature vector to be able to
analyze the signal for defects. How do we do this?

First, since an FM signal is made up of various components, one of the primary ways to analyze
it is to examine, in the frequency domain, its MPX (Stereo Multiplexed signal) spectrum plot.
Theoretically, it looks like what is shown in Figure 2 [15]. It is made up of four primary
components - the mono, pilot, stereo and RDS (Radio Data System) [16,17,18] components. The
mono and stereo parts will move up and down a bit according to the audio that is being
transmitted whereas the pilot signal is, theoretically, a constant value and serves, as its name
indicates, as a pilot or reference point. The RDS component contains the information that shows
up on the radio screen. For example, the name of the radio station, the song title, etc.

Figure 2. Theoretical MPX spectrum plot

71
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
The 10-second sample from the RTL-SDR device will produce a real-time representation of this
theoretical plot. Figure 3 shows what it might look like in real-life. We call this the “raw”
spectrum plot.

Figure 3. Real-time MPX spectrum plot (x-axis in MHz)

Several potential feature-vector variables can immediately be calculated from this empirical plot,
namely, the maximum value of the four primary components of the spectrum plot, their
respective powers, their relative powers with respect to the entire spectrum, and their relative
powers with respect to the total power in the rds spectrum interval. For example, the variable
“pilot” is the maximum value of the spectrum plot in the interval (0.018,0.20). Pilot_power is the
sum of the power in that interval, pilot_power0 is its percentage of the total power, and
pilot_power1 is the ratio of pilot_power and rds_power. Likewise, similar variables are
calculated for DB, stereo and rds using the vertical dashed line intervals in Figure 4. See also
Table 2.

Table 2. Spectrum plot feature vector variables

From IQ data DB pilot stereo rds

From wave file DB0 pilot0 stereo0 rds0
Freq intervals (0,.05) (.018,.020) (.037,.039) (.055,.059)

Likewise, using the IQ data file, a wave file of the 10-second audio clip was reconstructed by
means of some complex mathematics. After normalizing this wave file, we reversed the
mathematics and, once again, created a modified MPX spectrum plot where we extracted the
modified variables DB0, pilot0, stereo0 and rds0.

Many other potential feature vector variables describing the audio signal were extracted from
each wave file, some (but not all) of which are listed below.

72
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
Table 3. Additional feature vector variables

H Total entropy of a time wave.,

M Acoustic index based on the median of the amplitude envelope.,
ac Acoustic complexity
ad Acoustic Diversity
ae Acoustic evenness
rms The root mean square.
crest1 The crest factor of an audio signal
zcr The zero-crossing rate
Z90AB The number of times the station code for Radio Logos (90AB) is found in the
RDS part of the signal.

The Mel-Frequency Cepstral Coefficients (MFCC) are often used in speech recognition [19],
even in identifying illegal broadcast stations [20]. In our situation for identifying defective
signals we used the wave file and extracted 12 MFCC which is a 998x12 matrix. For each of the
12 columns, we calculated five statistics – the mean, the variance, the skewness, the kurtosis, and
after discretizing the vector of values into bins, the information entropy. For example, from the
1st MFC coefficient we extracted MC1, MC2, MC3, MC4 and MC5 which represent the above
five statistics. In the end, we had a vector of length 60.

These site/time variables were also added to the feature vector and after adding the interaction
term for Hour by Minute, the dimension of our final feature vector was over 150.

Table 4. Site/Time feature vector variables

Site The transmission site

Weekday Which day of the week.
Hour Which hour of the day
Minute (0 or 1) 0=1st half or hour, 1=2nd half.

The last step of our data preparation was to determine the defect status for every wave file. While
the norm would be to decide if the audio was either good or bad [21], we expanded the output
dimension to three. That is, the author of this paper listened carefully to every audio file and
decided if it was good (OK), bad (DF), or very bad (VB). The guideline that was followed was
that the audio would be labeled good if it sounded normal and bad if it contained some
background noise, a random pop noise or a minor distortion, something that a listener may or
may not notice. The audio was labeled very bad if it contained excess static beyond a background
noise, abnormal silence, major distortion, or was completely missing.

Although the problem is multi-nominal since we have three outcomes (OK, DF, VB), we divided
the problem into three binomial comparisons, namely, OK versus DF, DF versus VB, and OK
versus VB and used the majority opinion method to predict the status of the audio file.

To do this, we followed an ancient saying which says, “In the mouth of two or three witnesses
shall every matter be established.” We applied this good advice by constructing two different
statistical models (a logistic regression model and a decision tree) for each of the above three
comparisons. For the logistic models we used a variety of proven statistical procedures (e.g.
sequential and partial p-values, R2, AIC, multicollinearity, error rates of the confusion matrix,
stepwise constructions, etc) to determine the “best” equations with the fewest feature vector
variables. problem. The author did experiment with other more complex statistical methods (e.g.
Neural Networks), but their analyses did not yield any significant improvements. Thus, the
simpler techniques (logistic regressions and decision trees) were chosen.
73
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
4.2. Comparison 1: OK versus DF

For this comparison we forced our logistic regression models to first include the “site/time”
variables. Thereafter, we investigated the inclusion of other feature vector variables. The reason
for this is that not every transmission site nor transmission times are the same. Some sites have
more natural background noise than others due to the close proximity of other RF (Radio
Frequency) transmission sources which affect the RTL-SDR device. Likewise, some days of the
week and some hours of the day have more talk shows which increases the likelihood of
background noise being heard.

The training sample breakdown for this comparison was 1029 good and 727 bad audio files. The
best logistic regression model for this comparison is below. Since we are comparing OK with DF,
we are calculating the probability of the audio being bad (DF) versus good (OK). The syntax we
use is L12 to indicate that this is the logistic probability model for groups 1 and 2. A similar
syntax is used for the other models.

Obviously,

74
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
The ANOVA table reveals that, in fact, all the site/time variables are very significant and should
be included in the model. For example, ignoring the intercept, Weekday and Site factors, when
Hour =18 and Minute =0 the coefficients add up to 3.3+0+0=3.3, whereas for Hour=18 and
Minute =1, the coefficients add up to 3.3+0.6-3.6=0.3. This means that the probability of a bad
audio is lower for the second half of Hour 18. Many other hours had similar results. The reason
for this is that we tend to broadcast our talk shows during the first half of every hour. The other
significant feature vector variables are MC6, MC1, MC3 and pilot. For example, according to the
model, when MC6 increases the probability of DF decreases.

Table 5. ANOVA table for OK vs DF

Source Df SS MS F Pr(>F)
Weekday 6 27.38 4.56 58.1 0
Hour 23 71.70 3.12 39.7 0
Minute 1 48.93 48.93 622.7 0
Site 4 3.26 0.82 10.38 0
HourxMinute 23 50.19 2.18 27.3 0
MC6 1 80.61 80.61 1026 0
MC1 1 6.64 6.64 84.6 0
MC3 1 3.89 3.89 49.5 0
pilot 1 0.34 0.34 4.34 0.04
Residuals 1694 133.07 0.08

Although we attempted to allow the site/time variables to be included, the decision tree algorithm
chose only MC2, MC6, and MC7 to form the tree (see Table 6).

Table 6. Decision Tree for OK vs DF

Node Variable Probability Cut-off value

1 MC2 <72.0773
2 MC6 <-1.41261
4 <leaf> 0.73770
5 <leaf> 0.04995
3 MC6 <-1.50283
6 <leaf> 0.97210
7 MC7 <40.5805
14 <leaf> 0.3590
15 <leaf> 0.75860

For example, if MC2=79, MC6=-1.0 and MC7=40, the tree returns 0.35900 which means that we
would say that the signal is good (OK). The above probabilities will be labeled as PT12(DF). In
addition, the error rates from the 2x2 confusion matrices of these two models were in the 7.5%
range. This is not surprising since at times it is difficult to discern if the audio is good or bad.

4.3. Comparison 2: DF versus VB

The training sample breakdown for this comparison was 727 bad and 123 very bad audio files.
However, variable Hour and the interaction term HourxMinutes could not be used due to
insufficient data. Nevertheless, the best logistic regression model for this comparison is below.
The ANOVA table reveals that, although the variables Weekday, Minute and Site are almost all
significant, the most influential variable (high SS value) in the entire model is rds_adj which is
the ratio of rds and zcr. The reason for the site/time variables having a smaller impact than the

75
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
OK vs DF model is that VB defects occur randomly, whereas DF defects tend to show up
(although not exclusively) during the first half of the hour.

Table 7. ANOVA table for DF vs VB

Source df SS MS F Pr(>F)
Monday 1 0.26 0.26 8.059 0.004
Saturday 1 0.47 0.47 14.406 0.000
Sunday 1 0.05 0.05 1.614 0.204
Thursday 1 0.48 0.48 14.769 0.000
Tuesday 1 0.39 0.39 12.033 0.000
Wednesday 1 0.01 0.01 0.417 0.518
minute0 1 3.98 3.98 123.360 0
site103_5 1 0.00 0.00 0.011 0.915
site88_6 1 1.86 1.86 57.588 0
site94_2 1 0.74 0.74 23.053 0
site96_3 1 2.37 2.37 73.357 0
rds_adj 1 58.08 58.08 1798.561 0
stereo_power0 1 3.46 3.46 107.134 0
MC5 1 6.08 6.08 188.250 0
Residuals 835 26.97 0.03

The logistic regression formula is shown below.

Obviously,

The decision tree algorithm chose rds0_adj (rds0/zcr), MC5 and MC16. The details are below.

76
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
Table 8. Decision Tree results for DF vs VB

Node Variable Probability Cut-off value

1 rds0_adj <-240.195
2 MC5 <1.68935
4 <leaf> 0.72723
5 <leaf> 0.00694
3 MC16 <25.3674
6 <leaf> 0.99090
7 <leaf> 0.11111

The above probabilities will be labeled as PT 23(VB). In addition, the error rates from the
2x2 confusion matrices of these two models were in the 1% range. In other words, their
discriminating power is high.

4.4. Comparison 3: OK versus VB

The training sample breakdown for this comparison was 1029 good and 123 very bad audio files.
Like the previous category, the variable Hour and the interaction term HourxMinutes could not
be used. The ANOVA table reveals that the most influential variable in the entire model rds0_adj
which is the ratio of rds0 and zcr.

Table 9. ANOVA table for OK vs VB

Source df SS MS F Pr(>F)
Monday 1 0.41 0.41 12.43 0
Saturday 1 4.72 4.72 141.9 0
Sunday 1 2.38 2.38 71.44 0
Thursday 1 0.20 0.20 5.91 0.015
Tuesday 1 0.36 0.36 10.70 0.001
Wednesday 1 0.09 0.09 2.84 0.09
minute0 1 0.91 0.91 27.4 0
site103_5 1 0.05 0.05 1.47 0
site88_6 1 0.96 0.96 28.8 0.915
site94_2 1 0.00 0.00 0.01 0
site96_3 1 0.75 0.75 22.4 0
rds0_adj 1 34.91 34.91 1048.8 0
stereo_power0 1 13.68 13.68 411.1 0
MC5 1 12.51 12.51 375.8 0
crest1 1 0.13 0.13 4.0 0.047
Residuals 1136 37.81 0.03

The best logistic regression model is shown below.

77
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024

Obviously,

The decision tree algorithm chose rds0_adj, pilot, MC5 and stereo_power0. The details are
below.

Table 10. Decision Tree results for OK vs VB

Node Variable Probability Cut-off

1 rds0_adj <-239.107
2 MC5 <1.54551
4 pilot <-46.0433
8 <leaf> 1.0
9 <leaf> 0.20
5 <leaf> 0.0058
3 stereo_power0 <0.0767323
6 <leaf> 1.0
7 <leaf> 0.40

The above probabilities will be labeled as P T13(DF). In addition, the error rates from the 2x2
confusion matrices of these two models were in the 1% range. In other words, their
discriminating power is high.

4.5. The Majority Opinion Algorithm

As previously mentioned, we will use the majority opinion method to determine if an audio file is
good, bad, or very bad. That is, for a given audio file we calculate the following probabilities
using the above six formulas. For example, for audio file 1, we have P L12=.94 which means we
would choose DF, for PL23=0, we would choose DF, and so on. For audio file 2, we have PT12=.76
which means we would choose DF, for P T13=0.01, we would choose OK. The column labeled
“Final Choice” is the majority opinion.

78
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
Table 11. Majority Opinion Example

File PL12 PL23 PL13 PT12 PT23 PT13 Final

Choice
1 0.94 0.00 1 0.97 0.01 0.01
Choice DF DF VB DF DF OK DF
2 0.93 0.79 1 0.76 0.01 0.01
Choice DF VB VB DF DF OK DF

Using this procedure on the test data, the results show that there are only 2 concerning errors. The
other errors simply reflect the difficulty of discerning between a good and bad audio file as the
differences might be very minor background noise, a “pop”, or something else that is of little
importance. Serious errors would be to say that the audio file is good when, in fact, it is very bad.
Or, to say it is very bad when it is good. Fortunately, we have no such errors which indicates that
these six formulas have high discrimination power.

Table 12. Majority Opinion for the test data

Actual/Predicted OK DF VB
OK 229 23 0
DF 19 166 0
VB 0 2 31

Investigating these two audio files we find that the first file had a slight excess silence at the
beginning of the file, but the rest of the file was of excellent quality. In fact, this file could have
been classified as bad instead of very bad. The second file had some static during the last 3
seconds, so really was a very bad file. All-in-all, our data analysis shows that our AI-based
algorithm is very effective for the real-time monitoring of our radio network signals.

4.6. Implementation

The practical implementation of this algorithm consisted of several steps. First, this algorithm
was invoked several times per hour. Secondly, if the algorithm said that the file is good, then no
action was taken. If results indicated that the audio is very bad, then the results were immediately
relayed back to the studio (via the internet) for further investigation. If the algorithm said that
audio is bad, then usually it is a momentary and isolated event. However, if this “DF” status
continued to appear with several sequential observations, then it was an early warning sign that
the signal is degrading and heading in the direction of VB. Such a situation would be
communicated to the studio and warrant further investigation.

5. CONCLUSIONS
While AI technology is sometimes clouded in mystery and awe, our hope is that the reader will
have obtained a better understanding and appreciation of this field after studying the details of
our multifaceted AI-driven radio network. Some AI techniques can be very simple and involve
only basic computer science and NLP operations while others are more complex. As our analysis
demonstrated, AI techniques do not necessarily have to involve formulae of thousands of
parameters. Sometimes the AI solution can be simple, yet effective, as was demonstrated by our
application where the dimension of our most complicated formula was less than 60.

Finally, our hope is that this article will encourage other radio networks to develop and
implement such AI methodologies.
79
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024

ACKNOWLEDGMENTS
The author thanks the staff of Radio Logos for their dedication and help in the implementation of
this project, namely, Eri Cobani, Morena Cobani, Arben Dinoshi and Nancy Sturm. Their voices
formed the basis of many of our AI applications.

REFERENCES

[1] “Futuri Launches RadioGPT, The World’s First AI-Driven Localized Radio Content”, (February
23, 2023) https://futurimedia.com/, Cleveland, Ohio.
[2] Hughes, Alex. (January 16, 2023). “ChatGPT: Everything you need to know about OpenAI’s
GPT-3 tool”, BBC Science Focus Magazine. https://www.sciencefocus.com/future-
technology/gpt-3/
[3] OpenAI, Achaim, J., Adler, A., et al, (2023) “GPT-4 Technical Report”,
https://api.semanticscholar.org/CorpusID: 257532815.
[4] Ding, M., "A Systematic Review on the Development of Speech Synthesis," (2023) 8th
International Conference on Computer and Communication Systems (ICCCS), Guangzhou, China,
2023, pp. 28-33, doi: 10.1109/ICCCS57501.2023.10150729.
[5] Vachhani, K. & Mallari, R.A. (2015) “Experimental study on wide band FM receiver using
GNURadio and RTL-SDR” International Conference on Advances in Computing,
Communications and Informatics, ICACCI.
[6] Reynders, B., Iyare, R.N., Rajendran, S., Volski, V., Vandenbosch, G.A.E. & S. Pollin, (2018)
“Using Cheap RTL-SDRs for Measuring Electrosmog”. 2018 IEEE Symposium on
Communications and Vehicular Technology (SCVT), 2018.
[7] Moncayo, H.I.R., Ortiz, L.A.S., & Cuervo, V.M.M., (2018) “Low-cost spectrum monitoring
system based on Dragonboard 410C and RTL-SDR 2832U dongle”, Scientia et Technica Año
XXII, Vol. 23, No. 01, marzo de 2018.
[8] Sierra, E.G. & Ramirez, G.A., (2015) “Low Cost SDR Spectrum Analyzer and Analog Radio
Receiver Using GNURadio, Raspberry Pi2 and SDR-RTL Dongle”, Available on ResearchGate.
[9] Santiago, I., Vidal-Beltran, S., & Martinez-Pinon, F., (2018) “Spectrum Analyzer by Software
Defined Radio” Proceedings of the 2018 International Conference Mechatronics, Electronics and
Automotive Engineering (ICMEAE), pp93–97.
[10] Chen, W.T., Chang, K.T. & Ko, C.P., (2016) “Spectrum monitoring for wireless TV and FM
broadcast using software-defined radio”, Multimedia. Tools and Applications. Vol 75, Issue 16,
pp. 9819-9836.
[11] Pidanic, J., Valenta, V., & Juryca, K., (2017) “The radio coverage monitoring by low-cost system
based on SDR”,59th International Symposium ELMAR Proceedings. Zadar: Croatian Society
Electronics in Marine - ELMAR, s. 123 - 127 s. ISBN 978-953-184-230-3.
[12] Ahmed, M., Sulaman, A. & Ali, Z... (2023) “Real Time ECG Monitoring and QPSK Broadcasting
using SDR”, Available on ResearchGate.
[13] Juhana, T. & Girianto, S., (2017) “An SDR-based Multistation FM Broadcasting Monitoring
System”, 11th International Conference on Telecommunication Systems Services and Applications
(TSSA). Publisher: IEEE.
[14] Klapuri, A. & Davy, M., (2007) “Signal processing methods for music transcription”, Springer
Science & Business Media. Chapter 5.
[15] Kim, J. & Parihk, K., (2016) “Distribution of the Analog or Digital FM Composite Multiplex
Signal across IP Networks”, GatesAir, Mason, Ohio. NAB Show April 2016.
[16] Roppel, Carson, (2009) “An FM/RDS (Radio Data System) Software Radio”, Proceedings of the
54th International Scientific Colloquium, 07-10 September 2009.
[17] Liang, G., EK, T. &, Joe, K., (2017) “Introduction to FM-Stereo-RDS Modulation”, Verigy.
[18] Ford, J. & Catron, G., (2011) “Simultaneous Digital Demodulation and RDS Extraction of FM
Radio Signals”, https://dx.doi.org/10.32597/honors/4/.
[19] Sharma, G., Umapathy, K. & Krishnan, S., (2019) “Trends in audio signal feature extraction
methods”; https://doi.org/10.1016/j.apacoust.2019.107020.
[20] Huang, M., Yang, D. R., Zhu, D., et al. (2020) “FM Broadcast Monitoring Using Artificial
Intelligence”, Radio Science, 55(4), DOI: 10.1029/2019RS006885

80
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.5, September 2024
[21] Kang, K.R., Rim, G., Ri, C.H. & Sturm, G., (2024) “Real-time Analysis System for monitoring
FM Signals Using RTL-SDR Dongle”, Master degree thesis in Electrical Engineering Faculty,
Pyongyang University of Science and Technology, Pyongyang, DPRK.

AUTHOR
George Sturm was born in the USA but lives in Albania. He received his Ph.D.
from The Ohio State University in 1979. His research interests include statistics,
mathematics, electric engineering, and radio broadcasting.

81