The study was conducted within the city of Nairobi, Kenya, in East Africa, covering an area of 14 km radius from the city center. Nairobi is geographically defined by a latitude of 1˚09' and 1˚27' South, a longitude of 35˚59' and 37˚57' East, covering an area of approximately 696 km². It is the political, economic, and cultural capital of Kenya and one of the most densely populated and rapidly urbanizing cities in East Africa. According to census records, Nairobi’s population grew from 2.1 million in 2009 to 4.4 million by 2019 [28] .

With the increase in the city’s population, there has also been a significant increase in the vehicle population, with the number of registered vehicles growing from approximately 500,000 in 2012 to over 1.2 million by 2022 [29] The vehicular mix is highly heterogeneous, ranging from bicycles, motorcycles, saloon cars, SUVs to PSVs, trucks, and buses, contributing to non-uniform and often excessive noise. Nairobi’s varied land use, spanning residential, commercial, industrial, and institutional zones, further amplifies the variability of noise exposure across locations, making Nairobi a compelling case study for RTN challenges in urban cities in Africa. To ensure a representative assessment of noise exposure across Nairobi city, 42 monitoring sites were selected across the 17 sub-counties, reflecting residential, commercial, institutional, and industrial land use typologies. Preliminary site reconnaissance and traffic flow observations were used to guide the selection of locations and ensure a balanced representation of all zones. Each site was geolocated using GPS and mapped using Geographic Information Systems (GIS) to visualize spatial distribution and support subsequent analysis, see Figure 1 . The table in Appendix 1 ( Table A1 ) summarizes each site along with the rationale for its selection.

A handheld Android Smartphone, Samsung Galaxy A12 Model SM-A127F/DS, was used to capture RTN through its built-in microphone, see Figure 2 .

The audio recordings captured by the smartphone’s default audio recording application (Voice Recorder) were saved in the MPEG-4 Audio format (M4A). M4A is a high-quality audio format that uses Advanced Audio Coding (AAC) to compress the audio, hence reducing its file size while still maintaining much of the original sound quality. It is commonly used in modern smartphones because it offers better sound fidelity than older formats like MPEG-1 Audio Layer 3 (MP3), especially at the same file size. Unlike MP3, which may discard important audio details, M4A offers a better balance between file size and quality, making it suitable for environmental noise measurements where accuracy is important. To ensure that the noise data captured was accurate and detailed, the smartphone recorded at a sampling rate of 48 kHz and 16-bit depth.

The sampling rate means that the audio was captured 48,000 times per second. This rate is widely used in audio applications and is high enough to capture all sounds within the range of human hearing, which is up to about 20 kHz. The bit depth of 16-bit defines how detailed each of those measurements is. A higher bit depth captures more detail, allowing for a wide dynamic range, thus both soft and loud traffic sounds were recorded with good accuracy. The noise data was also captured at a bit rate of 256 kbps, meaning 256,000 bits of audio data were stored every second. This is considered high-quality audio, almost indistinguishable from lossless formats in many cases, and is well above the threshold needed to detect subtle sound differences, especially in noisy environments.

In parallel, a calibrated portable Class 1 Sound Level Meter (Lutron SD4033), conforming to IEC 61672-1 standards, see Figure 3 , was used to record noise levels. The SLM was set in fast response mode, which is ideal for recording the quickly varying nature of RTN. To account for the human ear’s response to sound, an A-weighting filter setting was used. High resolution was ensured by logging data at 2-second intervals during each 15-minute sampling session, enabling the computation of key acoustic metrics such as the L_eq.

At each of the 42 selected sites, noise data was recorded for 15-minute intervals every hour, starting from 6:00 A.M. to 6:00 P.M., resulting in 13 recordings per site, per day. For each day, the equivalent continuous sound levels (L_eq) collected were averaged to obtain a daily average L_eq. This process was repeated for 7 days, after which the final L_eq for each location was computed as the average of the 7 daily L_eq values. During data collection, both the SLM and the smartphone were handheld and positioned at an approximate height of 1.2 meters, maintaining a horizontal distance of about 1.5 meters from the road edge to minimize reflections and direct influence from passing vehicles. They were oriented towards the traffic flow to ensure direct exposure to incident sound waves, see Figure 4 and Figure 5 .

Once the audio recordings were collected using the smartphone, the audio M4A files were transferred to a computer for analysis. Rather than relying on application-displayed sound levels, the actual audio data was analyzed directly using Python, a powerful programming language. Python provided flexibility and control through libraries such as librosa and scipy.io for reading and analyzing the audio files, numpy for mathematical operations, and pandas for organizing and processing large volumes of data across multiple sites and times.

The M4A format, while efficient and storage-friendly, is not ideal for detailed acoustic analysis, as compression can lead to minor losses in signal accuracy. Therefore, for precise signal processing, the audio was converted to Waveform Audio File Format (WAV), an uncompressed, high-fidelity format commonly used in professional sound engineering and scientific research. WAV files preserve the quality and frequency content of the original signal, making them ideal for sound pressure level (SPL) computation. Each 15-minute recording was split into frames of 6-second duration to allow for frame-by-frame analysis of the sound pressure level (SPL). An A-weighting filter was first applied to mimic how the human ear responds to sound. Frequencies below 20 Hz or above 20 kHz are attenuated, and mid-frequency sounds between 1 kHz to 6 kHz are emphasized, aligning results with standard environmental noise assessments in dB(A). For each frame, the Root Mean Square (RMS) of the audio data was calculated using the formula:

$\text{RMS}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{{\displaystyle \sum }}}{\left(x_i\right)}^2}$(1)

where: $x_i$ is the amplitude of the i^th audio sample in the frame, N is the total number of samples in that frame, and RMS is the measure of the effective (average power) amplitude of the signal. To prevent computational issues caused by extremely small RMS values, which would result in negative infinity (−∞) as the result, a minimum threshold was applied to the RMS values using a constant (epsilon). After applying the threshold, the RMS values were converted into sound pressure level values in dB(A) using the standard logarithmic formula:

$L_i=20\times {\log }_{10}\left(\frac{\text{RMS}\left(t\right)}{p_0}\right)$(2)

where: $\text{RMS}\left(t\right)$ is the Root Mean Square amplitude of the audio signal at time t, $p_0=20\text{μPa}$ is the reference sound pressure in air and $L_i$ represents each instantaneous Sound Pressure Level value.

The Equivalent Continuous Sound Level (L_eq) for each 15-minute sampling session was then calculated as the energy-averaged SPL over all frames using the formula:

$L_{eq}=10{\log }_{10}\left(\frac{1}{N}\underset{i=1}{\overset{N}{{\displaystyle \sum }}}{10}^{\frac{L_i}{10}}\right)$(3)

where $Li$ represents each instantaneous SPL reading, and N is the number of 6-second frames in the 15-minute sample.

The SLM was on the other hand, was configured to log 2-second SPL values directly into an Excel spreadsheet throughout each 15-minute interval. The L_eq was computed on Excel using the logarithmic energy averaging formula as shown in Equation (3).

For both devices, the SLM and the smartphone, the hourly L_eq values were computed for each location. For each day, the equivalent continuous sound levels (L_eq) collected were averaged to obtain a daily average L_eq per location. This process was repeated for all the data, and finally, the L_eq for each location was computed as the average of the 7 daily L_eq values.

Previous studies have used various statistical methods to assess the accuracy and reliability of smartphone-based noise measurements when compared to standard Sound Level Meters. Kardous and Shaw (2014) evaluated several noise applications by computing the Mean Error between the smartphone-recorded levels and a calibrated Type 1 SLM using the formula:

$\text{Mean Error}=\frac{1}{N}\underset{i=1}{\overset{N}{{\displaystyle \sum }}}\left(L_{\text{smartphone}}-L_{\text{SPL}}\right)$ (4)

where N is the number of paired measurements. Their findings showed that certain apps revealed a Mean Error of ±2 dB under controlled indoor conditions.

D’Hondt et al. (2013) applied linear regression to quantify the relationship between smartphone and SLM readings, using the formula:

${\text{SPL}}_{\text{SLM}}=a\times {\text{SPL}}_{\text{smartphone}}+b$ (5)

where a and b are regression coefficients, and the strength of the relationship was assessed using the coefficient of determination (R²). R² values above 0.9 were reported for well-conditioned recordings.

Murphy and King, (2016) on their part used the Root Mean Square Error (RMSE) to evaluate accuracy, given by:

$\text{RMSE}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{{\displaystyle \sum }}}{\left(L_{\text{smartphone}}-L_i\right)}^2}$ (6)

and they also constructed Bland-Altman plots to visualize the bias and limits of agreement.

In this study, the Margin of Error was computed to quantify the level of uncertainty and assess the accuracy and viability of smartphones as tools for noise measurements compared to that of the reference SLM. By calculating the Margin of Error, the study established the confidence level using average smartphone readings across different locations, ensuring the scientific integrity of results derived from non-calibrated consumer-grade devices, and validating the use of smartphones as a potential and alternative tool for RTN measurement.

The Mean Absolute Error (MAE) was calculated using the formula:

$\text{Mean Absolute Error}=\frac{1}{N}\underset{i=1}{\overset{N}{{\displaystyle \sum }}}\left|\left(L_{\text{smartphone}}-L_{\text{SPL}}\right)\right|$ (7)

And the standard deviation is calculated from the formula:

$\sigma =\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{{\displaystyle \sum }}}{\left(x_i-\bar{x}\right)}^2}$ (8)

where $x_i$ is the error at each site $i$, $\bar{x}$ is the mean error and $N$ is the number of measurement sites. From these, the Margin of Error (MoE) at a 99% confidence level was computed using:

$\text{MoE}=z\times \frac{\sigma }{\sqrt{N}}$ (9)

where $z=2.58$ for 99% confidence and $\sigma$ is the standard deviation computed from Equation (9).

Given observed discrepancies between L_eq values measured by smartphones and those recorded using reference-grade SLMs, a linear regression model was developed, establishing the best line of fit hence calibrating the smartphone readings. This model allowed for adjustment of systemic deviations and assessment of whether smartphones could reliably be used as tools of RTN measurement by approximating SLM values. The regression analysis was conducted using Python, and the model aimed to express the SLM-measured L_eq as a function of smartphone-measured L_eq using the form:

$L_{eq\left(\text{SLM}\right)}={\beta }_0+{\beta }_1\cdot L_{eq\left(\text{smartphone}\right)}+\epsilon$(10)

where ${\beta }_0\text{and}{\beta }_1$ are regression coefficients and ε is a random error term.