Applied Acoustics 143 (2019) 222–228

Contents lists available at ScienceDirect

Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust

An averaging method for accurately calibrating smartphone

microphones for environmental noise measurement
Saurabh Garg ⇑, Kian Meng Lim, Heow Pueh Lee
Department of Mechanical Engineering, National University of Singapore, Singapore

a r t i c l e i n f o a b s t r a c t

Article history: Smartphones have become ubiquitous not only as communication devices but also as smart devices for
Received 9 May 2018 majority of the population. The small form-factor of smartphones together with a rich set of application
Received in revised form 27 July 2018 programming interfaces, low-power consumption, and presence of several types of sensors such as GPS,
Accepted 14 August 2018
microphone, gyroscope, accelerometer, barometer etc. makes them a good candidate for building apps for
Available online 27 September 2018
measuring and monitoring various environmental parameters. However, for these apps to be useful the
sensors must be first calibrated. In this paper, we focus on methodologies for accurately measuring sound
Keywords:
pressure level and frequency spectrum using microphone built into smartphones. For this purpose, we
Smartphone
App
present an averaging method for accurately calibrating a typical smartphone microphone against a ref-
Calibration erence microphone. We show experimental results illustrating that the proposed method can achieve
Sound level meters an accuracy of ±0.7 dB for 99.7% of measurements for three Samsung smartphones. We also present
results showing that it is possible to calibrate a smartphone using another smartphone calibrated using
our method.
Ó 2018 Elsevier Ltd. All rights reserved.

1. Introduction set of application programming interfaces provided by the smart-

phone can be used to build apps for measuring sound pressure
Noise pollution has become a major problem for many cities level [3], hearing-aid [4], coronary heart disease detector [5], earth-
due to increasing urbanization. This has resulted in increase of quake detector [6], and pathological tremor detector [7]. For these
exposure to noise (unwanted, unpleasant, often loud sound) for apps to accurately process audio data it is necessary to calibrate
people living in cities. Continuous exposure to high level of noise the smartphone microphone to report not only correct sound pres-
can cause both physiological and psychological problems such as sure levels but also correct frequency spectrum.
hearing impairment, hypertension, ischemic heart disease, annoy- Smartphones have become popular only in the last 10 years and
ance, and sleep disturbance [1]. Children living in noisy places have as such there are only few reported works [8–14] on computing
been shown to exhibit poor academic performance [2]. Low birth sound pressure level on smartphones. These works focus on urban
weight has been associated with exposure to loud noises in preg- noise mapping and use a single calibration value to correct the
nant women [2]. Thus, there is a strong need to not only educate sound pressure level on smartphones. The calibration value is com-
people about the adverse effects of the noise pollution but also puted in laboratory by comparing smartphone response with a ref-
to equip them with tools to monitor their exposure to noise. erence microphone using either white or pink noise. The
Sound level meter (SLM) is the standard tool to measure sound calibration techniques reported in these papers are sufficient for
pressure level (SPL) in dB or dBA (A-weighted to account for low mapping urban noise, however, they are not enough for many
sensitivity of human ear to low frequencies). However, these other applications such as psychoacoustics and determining fre-
meters are expensive, delicate, bulky and more importantly quency spectrum of a noise source to build effective noise barriers.
difficult-to-use by non-professionals. With the advent of smart- To make smartphones more versatile for applications which
phones, we all carry a powerful minicomputer equipped with a require audio processing, we propose that the calibration tech-
variety of sensors such as microphone, accelerometer, gyroscope, nique should have the following three characteristics:
GPS, barometer, and others. These sensors together with a rich Complete: calibration technique should not only calibrate for
sound pressure level but also for frequency spectrum as the latter
is useful in many applications such as psycho-acoustic analysis.
⇑ Corresponding author.
E-mail address: [email protected] (S. Garg).

https://doi.org/10.1016/j.apacoust.2018.08.013
0003-682X/Ó 2018 Elsevier Ltd. All rights reserved.
 S. Garg et al. / Applied Acoustics 143 (2019) 222–228 223

Cost-effective: calibration technique should be cheap to make reported that the errors in noise levels can vary from 1% to 12%
method accessible to large number of people and it should be fast depending on the value of these two parameters.
to allow calibrating a large number smartphones necessary for Rajib et al. [12] presented a context aware technique for noise
large scale participatory sensing experiments. mapping using smartphones. They built classifiers to detect the
Well-tested: The quality of microphone varies widely among location (hand, pocket, or bag) of phone and automatically deter-
various models of smartphones and it is possible that the calibra- mine when to start measuring. They used an in situ calibration
tion accuracy might differ among different smartphone of the same technique, where phone plays a reference sound and at the same
model. Therefore, it is important to sufficiently test calibration for time record it. The reference sound is compared with the recording
each model and include accuracy information along with the cali- to determine calibration factors. It is, however, not clear how this
bration data. technique will work as audio level of the sound played on phone
In this paper, we present an averaging method for accurately depends on several factors such as volume control, equalizers,
calibrating smartphone microphone against a reference class 1 phone casing, etc. Their measurement accuracy varies from 2.5 to
microphone. The proposed method calibrates smartphone micro- 4.91 dBA depending on the context when compared with sound
phone not only for sound pressure level but also for the frequency pressure level measured with a class 2 sound level meter.
spectrum. Calibration and validation were both done in the field, Navarrete et al. [3] presented a design of sound level meter.
thus expensive equipment such as anechoic or reverberation They used a single offset to calibrate the sound level meter against
chamber is not required. Currently, there are two main smart- a reference microphone and can achieve an accuracy of class 3
phone platforms: iOS from Apple and Android from Google. Some sound level meters (±3 dB). There are few other reported works
existing works [15–17] have reported that iOS devices with avail- Aumon et al. [14], NoiseCo [11], NoiseMap [8], and NoiseSpy [20]
able apps are more accurate compared to Android devices for noise on noise mapping which use single overall calibration factor. The
measurements. Although, our method can work equally well with calibration method in these works are similar and the error is
either iOS or Android devices we chose to test our method only on either not reported or is 1.5 dB or more.
Android devices to fill this gap. We will consider iOS devices in the There is a plethora of apps on both Google Play Store (Android)
near future. and Apple App Store (iOS). Most of the apps available comes with
either no calibration or allow a user to specify a single offset to
2. Related works adjust the reported sound pressure level. There are some reported
works which compare the accuracy of sound level meter apps for
One of the first comprehensive work on calibration and noise both Android and iOS [15–17]. These studies reported that iOS apps
mapping using smartphones is NoiseTube [9,18]. They used linear are superior in accuracy compared to Android apps. One potential
mapping between noise levels from phones and a reference micro- reason for this discrepancy is that there are only few variations of
phone for calibration. However, we have found that the response of iPhone in terms of hardware but there are many variations in
a smartphone microphone is linear with respect to noise level but devices when it comes to Android based smartphones. Kardous
non-linear with respect to frequency. They reported an accuracy of and Shaw [21] found that using an external calibrated microphone
0.56 dBA for noise levels under 100 dB in laboratory. In the field, improves the accuracy of sound level meter apps to within ±1 dB of
the accuracy varies as expected and for an 81 min walk the mean reference. In [22], Celestina et al. presented calibration method
error was 0.15 dB. For a 4-min interval during the walk the error used by NoiSee app available for iOS devices. They showed that
was 2.37 dB. In the presence of wind, the error reported was as NoiSee app running on iOS devices with MicW type i436 external
high as 10 dB. microphone can achieve compliance with most of the requirements
Zuo et al. [13] presented a technique for calibrating smart- for Class 2 of IEC 61672/ANSI S1.4-2014 standard.
phones using 1 kHz pure tone with varying amplitude to create
distinctive features for automatic comparison of signals from
SLM and smartphone. They did the experiment in an anechoic 3. Measuring noise levels on smartphone
chamber and compared the sound pressure level computed by
smartphone with the sound pressure level measured from a class A smartphone captures audio from its microphone and digitizes it
2 sound level meter. They determine a single overall calibration to an audio buffer consisting of floating-point numbers proportional
factor as a mean of sound pressure level differences between to the input sound pressure. To use a smartphone as a sound level
smartphone and reference microphone. They reported variable meter, an app running on the smartphone has to compute sound
accuracy among smartphones tested with accuracy under 1 dB pressure level and frequency spectrum from the input audio buffer.
for HTC Butterfly, Samsung S3, and Unistrong J4; and accuracy over According to the current international standard for sound level
2 dB for iPhone 6S and a second Unistrong J4. These results are meter, IEC 61672-1:2013 [23], sound level meter should also
based on only 10 measurements with unknown lengths. It seems include support for A- and C-frequency weighting. A-weighting
that all five phones were tested at the same time using a single ref- compensates for the fact that human ear is less sensitive to low fre-
erence microphone. The large distance between reference micro- quencies. A-weighted Leq or LAeq is reported in dBA and is com-
phone and smartphone microphone will affect the error between monly used for measuring environmental noise. The response of
two instruments due to directional sound, reflections, etc. It might the human ear depends not only on frequency but also the noise
also be the reason for large difference in error (0.59 vs 2.71 dB) level. At high noise levels (>100 dB), the ear’s response is flatter
between two models of Unistrong J4. In contrast, we measure with and is represented by C-weighting. C-weighted Leq or LC eq is
only one smartphone at the time and keep the reference micro- reported in dBC. C-weighting is frequently used for signals with
phone and smartphone microphone aligned. significant low frequency content, such as airport noise.
Zamora et al. [19] presented and compared accuracy of three A- and C-frequency weighting can be implemented in the time-
noise calculation algorithms with respect to the sampling fre- domain using digital filters or in the frequency-domain. Since our
quency and block size (length of measurement). Their method does goal is to compute not only sound pressure level but also frequency
not calibrate phone per se but determine the values of the sam- spectrum we have implemented frequency weighting in the fre-
pling frequency and block size which will produce the least error quency domain as specified in IEC 61672-1:2013 [23]. Fig. 1 illus-
in noise measurements. There are two side effects of this approach: trates the process of calculating sound pressure level and
loosing high-frequencies and lower frequency resolution. They frequency spectrum from the audio buffer:
224 S. Garg et al. / Applied Acoustics 143 (2019) 222–228

Fig. 1. The process of calculating sound pressure level and frequency spectrum
from audio buffer.
Fig. 2. Samsung S7’s internal microphone response is non-linear with respect to
frequency.
1. Use fast Fourier transform (FFT) to compute frequency spec-
trum from the input audio buffer.
2. Since the microphone in a smartphone is not calibrated we need for frequency spectrum. We do not require specialized scientific
to apply a calibration factor for each frequency to get a cali- equipment such as an anechoic chamber, signal generator, or pro-
brated frequency spectrum. The calibrated spectrum is a set of fessional speaker system. The calibration is done in the field only
pairs ðf n ; cðf n ÞjX½njÞ, where n 2 ½1; N=2. f n is the nth frequency with the help of a class 1 microphone connected to a data record-
in the spectrum. cðf n Þ and X½n are, respectively, the correction ing system.
factor and the coefficient in the Fourier transform of the input To calibrate, first we simultaneously record environmental
audio buffer for f n . N is the number of samples in the input noise in field for 2–5 min from a smartphone microphone and a
audio buffer. Sampling frequency f s and N determine the fre- class 1 microphone as uncompressed WAV files. Next, our calibra-
quency resolution ðf s =NÞ of the spectrum and the response time tion method computes the average frequency spectrum with fre-
ðN=F s Þ of the sound level meter. quency resolution of 1 Hz for both smartphone and reference
3. Optionally, add the frequency weighting to the spectrum using WAV files. The average spectrum is computed by dividing the time
the weighting function aW , where W is either A, C, or Z (no- signal into a set of overlapping frames and computing average of
weighting). The calibrated and frequency weighted spectrum FFT for all frames. We used frames with number of samples equal
is a set of pairs ðf n ; cðf n ÞaW ðf n ÞjX½njÞ, where n 2 ½1; N=2. to sampling frequency to obtain 1 Hz frequency resolution and 50%
4. Finally, the energy in the calibrated frequency spectrum is overlap between frames.
summed up to provide sound pressure level in dB, dBA, or The two spectrums, denoted by Sr (reference) and Sp (smart-
dBC, depending on the user’s choice not to apply any weighting phone) are defined in Eqs. (2) and (3).
or use either A- or C-weighting:   
2 2
Sr ¼ sri jsri ¼ 10log10 jX½ij ð2Þ
2 X N2
N=2
LW eq ¼ 10log10 2 cðf n ÞaW ðf n ÞjX½nj2 ð1Þ
N n¼1   
2
Sp ¼ spi jspi ¼ 10log10 2
jY½ij2 ð3Þ
N
4. Calibration method
where i 2 ½1; N=2 and N is the number of samples as well as sam-
The noise level computed directly from the audio data received pling frequency. X½i and Y½i are the ith coefficient in the frequency
by the smartphone microphone has its own frame of reference and spectrum for the reference and smartphone, respectively. The dif-
is not correct. To obtain the correct noise level, smartphone micro- ference between two spectrums are the correction factors required
phone must be calibrated against a standard reference micro- to calibrate smartphone microphone. The ith correction factor is
phone. Existing works have reported two methods for calibrating defined in Eq. (4).
smartphone microphones: single overall calibration factor [8,12],     !
and a set of calibration factors for different noise levels [9,18]. 2 2 2 2 jX i j2
ci ¼ 10log10 jX i j  10log10 jY i j ¼ 10log10 ð4Þ
Single calibration factor can be used when response of a micro- N2 N2 jY i j2
phone is linear with respect to both frequency and noise levels as is
Environmental noise in field is transient and may capture only a
the case with reference microphones. A set of calibration factors for
certain range of frequency at any given location therefore we
different noise levels are required if the response of the smart-
repeat the above process at several locations and for each fre-
phone microphone is not linear with respect to noise levels, which
quency we compute the correction factor as an average of correc-
was the case with older phone models [9,18]. However, we have
tion factors from all locations. We call the set fci ji 2 ½1; N=2g
found that newer smartphones running Android, such as Samsung
calibration table for the smartphone. Due to the central limit the-
S4 and newer, and ZTE Blade Q Pro have a linear response with
orem, the average correction factors will be close to the actual cor-
respect to noise levels up to 100 dB but the microphone response
rection factors if the number of samples is large.
is not linear with respect to frequency. Fig. 2 illustrates non-
linearity as the difference in sound pressure level between refer-
ence class 1 microphone and Samsung S7 microphone. 5. Materials and methods
In this paper, we propose a novel per-frequency method for
accurately calibrating smartphone microphone for environmental In this section, we first discuss all the software and hardware
noise mapping. Our method calibrates smartphone microphone tools we have used, followed by the experimental setup used to
for not only equivalent continuous sound pressure level but also calibrate smartphone microphones.
 S. Garg et al. / Applied Acoustics 143 (2019) 222–228 225

5.1. Materials We used PCB 377B02 class 1 microphone with B&K Sonoscout
Type 3663 data recording system as a reference, however, any sys-
We have developed a sound level meter app for Android called tem which can record audio as an uncompressed PCM WAV file can
NoiseExplorer. Fig. 3 illustrates the user interface of NoiseExplorer. be used.
The app is minimal with a single button to start/stop SLM. Once Next, we need a tool to compute average frequency spectrum
start is pressed, the app shows A-weighted instantaneous SPL from recorded WAV files. We developed a desktop software called
ðLAinst Þ on the left, and minimum ðLAmin Þ, average ðLAeq Þ, and max- NoiseMeter for this task. We used Java so that NoiseExplorer and
imum SPL ðLAmax Þ on the right. The live frequency spectrum is NoiseMeter can share same code (Android apps must be written
shown in the bottom half. The instantaneous SPL and frequency in Java). NoiseMeter is a command line application which accepts
spectrum is computed from a number of samples defined by the WAV files from smartphones as well as reference microphone
response type (35 ms, 125 ms, and 1 s) in settings. The calibration and outputs average frequency spectrum, spectrogram, and SPL
file when copied to the NoiseExplorer/Calibration directory in the (LWeq and LW max ). It provides the user with options to choose
phone, can be selected from the settings as shown in Fig. 3c. the amount of overlap between adjacent frames (0.0–1.0),
NoiseExplorer records audio from VOICE_RECOGNITION audio response type (35 ms, 125 ms, and 1 s), frequency weighting (Z,
source defined in Android. For VOICE_RECOGNITION audio source, A, and C), and calibration file. Our calibration method is imple-
manufactures of Android phones, are required to not use gain or mented as a Python script which uses NoiseMeter to compute
noise suppression pre-processing effects. After the app receives average frequency spectrum.
an audio buffer, it writes it to an uncompressed 16-bit integer We have tested our calibration method with four different
PCM WAV file and also accumulate samples as defined by the Android smartphones: two Samsung Galaxy S4 phones (both old
response type. Once the required number of samples are collected, and used for more than 3 years), one Samsung Galaxy S7 (rela-
the app follows the process defined in Fig. 1 to compute LAinst and tively new and used for about 2 years), and one ZTE Blade Q Pro
frequency spectrum. If the calibration factors, as defined in the (new low-cost phone about 60USD). We used Samsung phones
calibration file, are not available for frequencies at which the because it is the largest manufacturer of Android based smart-
frequency spectrum is computed, the app linearly interpolates phones. Once the proposed calibration method was established
calibration factors for two adjoining frequencies in the and validated using Samsung phones we decided to acquire a
calibration file. new low-cost phone. We chose ZTE Blade Q Pro because it was

Fig. 3. NoiseExplorer user interface. (a) Main screen, (b) settings screen, and (c) selecting calibration file.

Fig. 4. The setup used for calibrating smartphone microphone.

226 S. Garg et al. / Applied Acoustics 143 (2019) 222–228

the cheapest phone available locally in Singapore at that time. The tables for them using procedure described in Section 4. The calibra-
second Samsung S4 phone had a protective casing on it. We tion tables for four phones are presented in Fig. 5. It can be seen
decided not to remove it since cases are integral part of how peo- that all four phones have a non-linear response with respect to
ple personalize their phones and effect of cases should be tested as the frequency. The figure also shows that the different phone mod-
well. els can have very different calibration tables. Calibration table is
similar for two Samsung S4 phones except for the dip around
5.2. Experimental setup 7 kHz, which is most likely because S4 #2 had a protective case
and S4 #1 was without a case.
To calibrate a smartphone, we first identify the opening of the Next, we computed sound pressure level ðLAeq Þ for all record-
microphone on the smartphone. Next, we hold the reference ings from both reference microphone and smartphones. LAeq was
microphone along the back side of the smartphone such that the computed using 1 Hz resolution and 50% overlap between frames.
center of two microphones are aligned and point in the same direc- Mean and standard deviation of the error (reference microphone –
tion as illustrated in Fig. 4. We do not use tape to mount reference smartphone) is presented in Table 2 and the distribution of error
microphone to the smartphone (tape seen in the figure is to tag the for all phones is presented in Fig. 6. Assuming error is distributed
microphone) and the two devices are hand-held at a comfortable normally, statistically speaking 99.7% of measurements have an
distance from the body (wall in the figure is used to create con- error within ±0.7 dB (mean ± 3  standard deviation) for three
trast; measurements were taken away from wall). All measure- Samsung phones. However, for the remaining 0.3% of measure-
ments were taken outdoors in open spaces while standing still ments error magnitude can be larger than 0.7 dB. In particular,
and consists of urban environmental noise such as machinery we have found that in the presence of wind noise error magnitude
(air conditioner compressors, air compressors, etc.), construction can be as large as 12 dB. Although, typically error magnitude is
(piling, excavating, drilling, hammering, etc.), moderate traffic between 2 and 5 dB.
(cars, buses, trucks, occasional airplane), people talking, bird calls, ZTE which is a low-cost device has a lower accuracy (±2.1 dB
insect noise, etc. for 99.7% of measurements) compared to other Samsung phones
which are much more expensive. In general, we expect
6. Experimental results cheaper devices to be less accurate than more expensive ones.
If necessary, accuracy of low-cost device can be improved by
In this section, we present four experiments illustrating the using an external microphone as has been reported by Kardous
effectiveness of our calibration method. In the first experiment and Shaw [21].
(Section 6.1) we discuss the accuracy achieved in computing Similar to calibration charts provided with reference micro-
sound pressure level using the proposed method. Since, the pro- phones, along with calibration factors we store the number of sam-
posed calibration method averages data from a number of mea- ples and their total length used to compute calibration table,
surements, in the second experiment (Section 6.2) we discuss number of measurements and their total length used to test the
the effect of sample size on the accuracy of computing sound pres- calibration file, and mean and standard deviation of error for these
sure level. In the third experiment (Section 6.3) we discuss the dif- measurements in the calibration file.
ferences in calibration tables of two Samsung S4 smartphones.
Finally, in the fourth experiment (Section 6.4) we show that it is 6.2. Effect of sample size on accuracy
possible to calibrate one smartphone using another calibrated
smartphone. Our calibration method computes the calibration table for a
For first three experiments, we recorded about 9 h of urban smartphone by averaging a set of differences (samples) in the fre-
environmental noise at 12 sites (182 recordings) in Singapore quency spectrum between a reference and smartphone micro-
using these phones between October 2016 and March 2017. The phone. Theoretically, the central limit theorem dictates that
recordings were made according to the calibration setup discussed averaging will produce results close to the actual calibration table
in Section 5.2. The details of the recordings are given in Table 1. if the number of samples are large. In this experiment, we investi-
The LAeq in 182 recordings ranged from 52.8 to 81.4 dBA, with gated the effect of sample size on the accuracy of computing sound
36% of measurements between 70 and 75 dBA, 30% between 65 pressure level for each phone.
and 70 dBA, 20% between 75 and 80 dBA, and 10% between 60 For Samsung Galaxy S4 #1 and Samsung S7 we randomly sam-
and 65 dBA. pled 5, 10, and 15 pairs of measurements and computed calibration
For the last experiment, we recorded about 25 min of data (14 table for each sample. We repeated this step three times for each
recordings) simultaneously from Samsung S4 #1, Samsung S4 #2, sample size resulting in total of nine calibration tables for each
and a reference microphone. phone. Similarly, for Samsung S4 #2 and ZTE Blade Q Pro we ran-
domly sampled 5 and 10 pairs of measurements three times each,
6.1. Calibration accuracy resulting in nine calibration tables for each phone. We choose
smaller number of samples for last two phones as they have fewer
For each of the four phones, we used five randomly sampled measurements and large sample size would naturally result in
recordings from all available recordings and computed calibration smaller error. It can be seen in Table 3 that even using five samples
is enough to achieve a low error and using more samples does not
lead to smaller errors. To find the most accurate calibration table,
we use few different sample sizes between five and ten and use
Table 1
Recording information for four smartphones. the one that give the smallest error.

Phone Recording time # Sites # Recordings

(hh:mm:ss) 6.3. Comparing smartphones of the same model
Samsung S4 #1 06:54:38 9 136
Samsung S4 #2 00:27:02 1 12 In this experiment, we investigated if the calibration table for
Samsung S7 00:43:52 1 19 one Samsung Galaxy S4 smartphone can be used for the other
ZTE Blade Q Pro 00:33:09 1 15
phone and vice versa.
 S. Garg et al. / Applied Acoustics 143 (2019) 222–228 227

Fig. 5. Calibration tables for four smartphones used in experiments.

Table 2 We used the calibration table corresponding to the first row in

Mean and standard deviation of error in computing A-weighted equivalent contin- Table 3 for S4 #2 for the recordings from S4 #1 and computed the
uous SPL ðLAeq Þ.
mean and standard deviation of error in computing sound pressure
Phone Mean (dB) Std. Dev. (dB) level and vice versa. The error for S4 #1 (using calibration table for
Samsung S4 #1 0.1 0.2 S4 #2) was 0.0 ± 0.3 dB and for S4 #2 (using calibration table for S4
Samsung S4 #2 0.1 0.2 #1) was 0.3 ± 0.2 dB. Although errors are higher in this case, as
Samsung S7 0.0 0.2 compared to when phones are calibrated against reference micro-
ZTE Blade Q Pro 0.2 0.7
phone, they are still small. This experiment illustrates that it might
be possible to use one calibration file for all smartphones of a par-
ticular model, however, more testing is required to validate this.
Since, phone cases are integral part of how people personalize their
phones effect of cases also needs to be tested.

6.4. Calibrating one smartphone using another calibrated smartphone

In this experiment, we investigated if we can use one calibrated

smartphone to calibrate another smartphone. Using the 14 record-
ings (25 min), we first calibrated Samsung S4 #1 with respected to
the reference microphone, and than calibrated Samsung S4 #2
with respected to Samsung S4 #1. The mean and standard devia-
tion of error in computing sound pressure level for Samsung S4
#2 is 0.2 ± 0.3 dB. We repeated the same experiment but first cal-
ibrated Samsung S4 #2 using the reference microphone and then
used it to calibrate Samsung S4 #1. The mean and standard devia-
tion of error in computing sound pressure level for Samsung S4 #1
is 0.2 ± 0.3 dB. These results show that our proposed method can
also be used to accurately calibrate one smartphone using another
Fig. 6. Distribution of error in computing sound pressure level. calibrated smartphone.

Table 3
Effect of sample size on sound pressure level accuracy. Numbers below are mean error (dB) in computing sound pressure level for all measurements for a phone ± standard
deviation of all errors.

Sample size Samsung S4 #1 Samsung S4 #2 Samsung S7 ZTE Blade Q Pro

5 0.1 ± 0.2 0.1 ± 0.2 +0.0 ± 0.2 0.2 ± 0.7
5 +0.0 ± 0.2 +0.1 ± 0.2 +0.2 ± 0.2 0.3 ± 0.7
5 0.2 ± 0.2 +0.1 ± 0.2 0.1 ± 0.2 0.5 ± 0.7
10 0.2 ± 0.2 +0.0 ± 0.2 +0.1 ± 0.2 0.3 ± 0.7
10 +0.0 ± 0.2 +0.0 ± 0.2 +0.1 ± 0.2 0.4 ± 0.7
10 +0.0 ± 0.2 0.1 ± 0.2 +0.1 ± 0.2 0.2 ± 0.7
15 +0.0 ± 0.2 – +0.0 ± 0.2 –
15 0.2 ± 0.2 – +0.0 ± 0.2 –
15 0.1 ± 0.2 – +0.0 ± 0.2 –
228 S. Garg et al. / Applied Acoustics 143 (2019) 222–228

7. Conclusions [5] Pareek V, Sharma RK. Coronary heart disease detection from voice analysis. In:
IEEE students’ conference on electrical, electronics and computer science
(SCEECS); 2016. p. 1–6.https://doi.org/10.1109/sceecs.2016.7509344.
In this paper, we presented an averaging method for accurately [6] Kong Q, Kwon YW, Schreier L, Allen S, Allen R, Strauss J. Smartphone-based
calibrating smartphone microphone against a reference class 1 networks for earthquake detection. In: 15th international conference on
innovations for community services (I4CS); 2015. p. 1–8.https://doi.org/10.
microphone. The proposed method can calibrate a smartphone
1109/i4cs.2015.7294490.
microphone with greater accuracy compared to the existing meth- [7] Daneault JF, Carignan B, Codère CÉ, Sadikot AF, Duval C. Using a smart phone as
ods. We showed that the proposed method has an accuracy of a standalone platform for detection and monitoring of pathological tremors.
Front Hum Neurosci 2012;6(357). https://doi.org/10.3389/fnhum.2012.00357.
0.7 dB for 99.7% of the measurements for three Samsung phones
[8] Schweizer I, Darmstadt T, Probst F, Bärtl R, Darmstadt T, Mühlhäuser M, et al.
used. This is a significant improvement over the existing reported Noisemap – real-time participatory noise maps. In: Second international
methods. However, for the remaining 0.3% of measurements error workshop on sensing applications on mobile phones; 2011.
magnitude can be larger than 0.7 dB. Especially, in the presence of [9] D’Hondt E, Stevens M, Jacobs A. Participatory noise mapping works! An
evaluation of participatory sensing as an alternative to standard techniques for
wind noise errors are unpredictable and we have observed errors environmental monitoring. Pervasive Mob Comput 2013;9(5):681–94. https://
as high as 12 dB. We plan to develop a wind noise filter to warn doi.org/10.1016/j.pmcj.2012.09.002.
users when noise levels are not accurate in the presence of wind. [10] Dumoulin R, Voix J. Calibration of smartphone-based devices for noise
exposure monitoring: methodology, uncertainties of measurement and
In this study, we have tested our method on three Samsung phones implementation. Proc Meet Acoust 2013;19(1):030037. https://doi.org/
and one low-cost ZTE phone. In the future, we plan to test our 10.1121/1.4800063.
method on devices from other popular Android phone manufactur- [11] Al-Saloul AHA, Li J, Bei Z, Zhu Y. Noiseco: Smartphone-based noise collection
and correction. In: 2015 4th international conference on computer science and
ers such as Huawei and Xiaomi. We also plan to test our method on network technology (ICCSNT), vol. 01; 2015. p. 369–73.https://doi.org/10.
iOS devices. 1109/iccsnt.2015.7490771.
The proposed method has three important characteristics [12] Rana R, Chou CT, Bulusu N, Kanhere S, Hu W. Ear-phone: a context-aware noise
mapping using smart phones. Pervasive Mob Comput 2015;17(A):1–22.
defined in this paper. It is complete (calibrate for both sound pres-
https://doi.org/10.1016/j.pmcj.2014.02.001.
sure level and frequency spectrum), cost-effective (does not [13] Zuo J, Xia H, Liu S, Qiao Y. Mapping urban environmental noise using
require special equipment such as anechoic chamber, noise gener- smartphones. Sensors 2016;16(10). https://doi.org/10.3390/s16101692.
[14] Aumond P, Lavandier C, Ribeiro C, Boix EG, Kambona K, D’Hondt E, et al. A
ator, or professional speaker system), and well-tested (tested with
study of the accuracy of mobile technology for measuring urban noise
9 h of environmental noise data). pollution in large scale participatory sensing campaigns. Appl Acoust
We showed that calibration file from one smartphone can be 2017;118:219–26. https://doi.org/10.1016/j.apacoust.2016.07.011.
used with another smartphone of same model, and we can cali- [15] Kardous CA, Shaw PB. Evaluation of smartphone sound measurement
applications. J Acoust Soc Am 2014;135(4):186–92. https://doi.org/10.1121/
brate one smartphone using another smartphone calibrated using 1.4865269.
the proposed method. However, more testing is required to vali- [16] Murphy E, King EA. Testing the accuracy of smartphones and sound level
date these findings. meter applications for measuring environmental noise. Appl Acoust
2016;106:16–22. https://doi.org/10.1016/j.apacoust.2015.12.012.
[17] Roberts B, Kardous C, Neitzel R. Improving the accuracy of smart devices to
Acknowledgments measure noise exposure. J Occup Environ Hyg 2016;13(11):840–6. https://doi.
org/10.1080/15459624.2016.1183014.
[18] Maisonneuve N, Stevens M, Niessen ME, Steels L. Noisetube: measuring and
This material is based on research supported by the Singapore mapping noise pollution with mobile phones. In: Information technologies in
Ministry of National Development and National Research Founda- environmental engineering: proceedings of the 4th international ICSC
tion under L2 NIC Award No. L2NICCFP1–2013–8. Saurabh Garg symposium Thessaloniki; 2009.https://doi.org/10.1007/978-3-540-88351-
716.
thank Shahrokh Sepehrirahnama at the National University of Sin-
[19] Zamora W, Calafate CT, Cano JC, Manzoni P. Accurate ambient noise
gapore for many valuable discussions and feedback on designing assessment using smartphones. Sensors 2017;17(4):917. https://doi.org/
experiments for the research project. 10.3390/s17040917.
[20] Kanjo E. Noisespy: a real-time mobile phone platform for urban noise
monitoring and mapping. Mob Netw Appl 2010;15(4):562–74. https://doi.
References org/10.1007/s11036-009-0217-y.
[21] Kardous CA, Shaw PB. Evaluation of smartphone sound measurement
[1] Stansfeld SA, Matheson MP. Noise pollution: non-auditory effects on health. applications (apps) using external microphones – a follow-up study. J Acoust
Brit Med Bull 2003;68:243–57. https://doi.org/10.1093/bmb/ldg033. Soc Am 2016;140(4):327–33. https://doi.org/10.1121/1.4964639.
[2] Kawada T. The effect of noise on the health of children. J Nippon Med Sch [22] Celestina M, Hrovat J, Kardous CA. Smartphone-based sound level
2004;71(1):5–10. https://doi.org/10.1272/jnms.71.5. measurement apps: evaluation of compliance with international sound level
[3] Navarrete R, Rocha MF, Salvador V. L, Orozco L, Calva CA, Osnaya MR, et al. meter standards. Appl Acoust 2018;139:119–28. https://doi.org/10.1016/j.
Software and hardware to replace a sound level meter. Int J Comput Commun apacoust.2018.04.011.
Eng 2012;1(4):297–300. https://doi.org/10.7763/ijcce.2012.v1.77. [23] IEC 61672-1:2013. Electroacoustics – sound level meters – Part 1:
[4] BioAid. Bioaid – the biologically inspired hearing aid. Online; 2016. URL:http:// Specifications. Standard. International Electrotechnical Commission; 2013.
bioaid.org.uk/.