Mobile Phone Identification from Recorded Speech Signals Using
Non-Speech Segments and Universal Background Model Adaptation
Dimitrios Kritsiolis
a
and Constantine Kotropoulos
b
Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki 54124, Greece
Keywords:
Mobile Phone Identification, Digital Speech Processing, Audio Forensics, Voice Activity Detection,
Universal Background Model.
Abstract:
Mobile phone identification from recorded speech signals is an audio forensic task that aims to establish the
authenticity of a speech recording. The typical methodology to address this problem is to extract features
from the entire signal, model the distribution of the features of each phone, and then perform classification
on the testing data. Here, we demonstrate that extracting features from non-speech segments or extracting
features from the entire recording and modeling them using a Universal Background Model (UBM) of speech
improves classification accuracy. The paper’s contribution is in the disclosure of experimental results on
two benchmark datasets, the MOBIPHONE and the CCNU Mobile datasets, demonstrating that non-speech
features and UBM modeling yield higher classification accuracy even under noisy recording conditions and
amplified speaker variability.
1 INTRODUCTION
Mobile phone identification from recorded speech
signals aims to establish the authenticity of a speech
recording. It is not uncommon for speech record-
ings from mobile phones to be presented as evidence
in criminal investigations or other legal proceedings.
However, for the recording to be admitted as evidence
in a legal case, the court must be convinced of its au-
thenticity and integrity (Maher, 2009).
Here, we examine the problem of brand and model
identification of mobile phones from recorded speech
signals. The term brand refers to the mobile phone
manufacturer, e.g. LG, and the term model refers to a
specific product from a manufacturer, e.g. LG L3.
A three-stage process can describe mobile phone
identification from recorded speech signals. In the
first stage, features capable of capturing the intrinsic
trace each device leaves on the speech signal are ex-
tracted. In the second stage, the extracted features
are used to train models that represent the distribu-
tion of features of each mobile phone. In the third
stage, classification is performed using the trained
models and the extracted features. The features to
be examined are the Mel Frequency Cepstral Coef-
a
https://orcid.org/0009-0009-3845-5928
b
https://orcid.org/0000-0001-9939-7930
ficients (MFCCs) (Davis and Mermelstein, 1980) ex-
tracted on a frame basis. The extracted feature vec-
tors are then used to train Gaussian Mixture Models
(GMMs) using the Expectation Maximization (EM)
algorithm (Dempster et al., 1977). The GMMs can
be used for Maximum Likelihood (ML) classification
or the extraction of fixed dimensional feature vectors,
known as Gaussian Supervectors (GSVs) (Garcia-
Romero and Espy-Wilson, 2010). The GSVs are ob-
tained by concatenating the mean vectors (and option-
ally the main diagonals of the covariance matrices) of
the GMMs. They are used in Support Vector Machine
(SVM) classification.
We will explore two different approaches to mo-
bile phone identification. The first approach is to
extract the features from the non-speech segments
of the signal since the speech parts might introduce
speaker-related noise into the device’s intrinsic trace.
The second approach stems from the fact that mo-
bile phone identification based on recorded speech
signals is similar to the speaker recognition prob-
lem. Thus, approaches used in the latter field can
be applied in the former. One of these approaches
is the Universal Background Model (UBM) of speech
(Reynolds et al., 2000) that is employed to adapt the
mobile phones’ feature vectors by means of Maxi-
mum A Posteriori (MAP) adaptation. GSVs are then
extracted from the UBM-adapted GMMs and are used
Kritsiolis, D. and Kotropoulos, C.
Mobile Phone Identification from Recorded Speech Signals Using Non-Speech Segments and Universal Background Model Adaptation.
DOI: 10.5220/0012420400003654
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 793-800
ISBN: 978-989-758-684-2; ISSN: 2184-4313
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
793

in the classification stage.
The paper contributes to disclosing experimental
evidence using either ML, SVM, or neural classifi-
cation on two benchmark datasets, namely the MO-
BIPHONE dataset and the CCNU Mobile dataset.
The empirical results demonstrate that both the non-
speech and the UBM approaches yield higher accu-
racy than the traditional approach of extracting fea-
tures from the entire speech signal and modeling them
with GMMs trained via the EM algorithm, even un-
der noisy recording conditions and amplified speaker
variability.
2 RELATED WORK
(Hanilci et al., 2011) proposed a mobile phone iden-
tification system using MFCCs, Vector Quantization
classifiers, and SVMs. (Hanilc¸i and Kinnunen, 2014)
extended their previous work by showing that MFCCs
obtained from the non-speech segments yield higher
mutual information than the MFCCs obtained from
the speech segments or the entire speech recording.
(Kotropoulos and Samaras, 2014) used MFCCs to
train GMMs and to obtain GSVs, which were used
with SVMs, a Radial Basis Function neural network,
and a Multilayer Perceptron. To test their methods,
they collected the MOBIPHONE dataset. (Kotropou-
los, 2014) obtained sketches of features from large-
size raw feature vectors and used them with a Sparse
Representation Classifier and SVMs on the Lincoln-
Labs Handset Database (Reynolds, 1997) and on the
MOBIPHONE dataset.
(Baldini and Amerini, 2019) proposed a mo-
bile phone identification approach by stimulating the
phones’ microphones with non-voice sounds. They
extracted the frequency spectrum magnitude and used
it as the input to a Convolutional Neural Network
(CNN), which performed the identification task. (Bal-
dini and Amerini, 2022) continued their work by ob-
taining spectral entropy features based on Shannon
entropy and Renyi entropy to implement dimension-
ality reduction of the spectral representation of audio
signals. These features were then the input to their
previous CNN.
(Giganti et al., 2022) proposed identifying mobile
phones from their microphone under noisy conditions
by applying convolutional neural network-based de-
noising on the signal’s spectrogram.
(Berdich et al., 2022) explored mobile phone mi-
crophone fingerprinting based on human speech, en-
vironmental sounds, and several live recordings per-
formed outdoors, using supervised machine learning
methods.
(Zeng et al., 2020) proposed an end-to-end deep
source recording device identification system for web
media forensics utilizing the MFCCs, the GSVs, and
the i-vectors. (Zeng et al., 2023) expanded their work
by proposing a spatio-temporal representation learn-
ing framework utilizing the MFCCs and the GSVs.
(Berdich et al., 2023) surveyed smartphone finger-
printing technologies based on sensor characteristics.
(Qamhan et al., 2023) proposed a transformer net-
work for authenticating the source microphone. The
Audio Forensic Dataset for Digital Multimedia Foren-
sics (Khan et al., 2018) and the King Saud University
speech database (Alsulaiman et al., 2013) were used.
(Leonzio et al., 2023) addressed audio splicing de-
tection and localization by using the CNN in (Baldini
and Amerini, 2019) for mobile phone classification.
They applied K-means to extracted features from in-
termediate layers of the CNN to detect splicing and
employed the cosine distance for localization between
consecutive frame vectors.
3 MOBILE PHONE
IDENTIFICATION
Due to tolerances in the manufacturing of electronic
components, each implementation of an electronic
circuit, particularly when a microphone is included,
cannot have the same transfer function (Hanilci et al.,
2011). The recorded speech signal’s frequency spec-
trum can be considered as the multiplication of the
spectrum of the original speech signal with the trans-
fer function of the mobile phone’s microphone. Con-
sequently, every mobile phone leaves its own device
fingerprint on the recorded speech signal.
Mobile phone identification is very similar to
closed-set text-independent speaker recognition. A
simple block diagram of an automatic speaker recog-
nition system is shown in Fig. 1. The same princi-
ples used in speaker recognition also apply to mobile
phone identification. Closed-set mobile phone identi-
fication is done in 3 stages: feature extraction, model-
ing, and classification.
Figure 1: Block diagram of a basic closed-set automatic
speaker recognition system.
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
794

The MFCCs are the features used and they are
based on the speech signal spectrum. The shape of
the vocal tract gives us information about the speaker
uttering the speech or the phoneme being uttered. It is
found in the spectral envelope of the short-time power
spectrum. The MFCCs try to represent that envelope.
Once the MFCCs have been extracted, each mo-
bile phone model or brand is modeled using a GMM
trained on the corresponding short-time features us-
ing the EM algorithm. The GMM probability den-
sity function is used to measure the similarity between
the testing data of an unknown mobile phone and the
trained GMMs. To obtain a fixed-dimensional vector
representing a speech utterance for SVM and neural
classification, we construct GSVs by concatenating
the parameters of a GMM trained on the speech ut-
terance.
After training the GMMs and extracting the
GSVs, classification is performed. Firstly, an ML
classifier is utilized by using the MFCC vectors ex-
tracted from the testing samples and comparing their
similarity with the trained GMMs. SVMs are also
trained using the GSVs extracted from the GMMs.
Finally, we employ the neural architecture in (Zeng
et al., 2023) for mobile phone identification that con-
siders both the MFCCs and the GSVs. The mobile
phone identification process outlined above will be re-
ferred to as the traditional approach. We examine two
different modifications to the traditional approach.
The first modification is referred to as the
non-speech approach. Speech parts have speaker-
dependent information that is irrelevant to the task at
hand. The noise-like signals in the non-speech parts
have a flatter spectral density and better capture the
mobile phone’s recording circuitry transfer function.
Accordingly, the feature extraction stage is modified
to extract MFCCs from non-speech parts. To obtain
non-speech parts, we use either a voice activity detec-
tion algorithm (Sohn et al., 1999) or speech enhance-
ment to estimate a clean speech signal and subtract it
from the initial speech signal to obtain a residual noise
signal. The speech enhancement method utilizes the
noise estimation method described in (Martin, 2001).
The second modification, referred to as the UBM
approach, stems from the fact that the problem of mo-
bile phone identification from recorded speech sig-
nals is very similar to the speaker recognition prob-
lem (Hanilci et al., 2011). A speech signal is mod-
eled in the time domain as the convolution of the vo-
cal cords’ excitation and the vocal tract’s impulse re-
sponse. Speaker recognition aims to extract features
that can distinguish the transfer function of the vo-
cal tract of each speaker. A recorded speech signal,
however, also contains a convolutional distortion in-
troduced by the mobile phone’s microphone. Sup-
pose that the recorded speech signal is produced by
a transfer function given by the serial composition of
the vocal tract transfer function and the microphone’s
transfer function. In that case, the task at hand be-
comes tantamount to speaker recognition. Thus, the
use of a UBM model, a large speaker-independent
GMM trained via the EM algorithm to capture the dis-
tribution of the MFCCs in a very large set of speak-
ers, becomes justified. The UBM is used to adapt the
GMMs directly via MAP adaptation.
4 DATASETS
4.1 MOBIPHONE and Augmentations
The MOBIPHONE dataset (Kotropoulos and Sama-
ras, 2014) consists of recorded speech utterances of
24 speakers, 12 males and 12 females, using 21 mo-
bile phones belonging to 7 brands. The 24 speakers
were randomly chosen from the TIMIT dataset (Garo-
folo et al., 1988). Each speaker uttered 10 sentences
approximately 3 seconds long. Each speaker’s sen-
tences were concatenated in a single WAV file. Thus,
the dataset contains 504 WAV files sampled at 16kHz,
one for each speaker recorded on each mobile phone.
To compare the performance of the different fea-
ture extraction and modeling approaches, 8 augmen-
tations were applied to the entire dataset to create 8
additional versions of the MOBIPHONE dataset.
The first set of 3 augmentations aims to simulate
different recording conditions. It includes the addi-
tion of Gaussian noise at a random signal-to-noise ra-
tio (SNR) between 10-20 dB, the addition of back-
ground chattering noise at a random SNR between
10-20 dB, and the addition of reverberation.
The second set of another 5 augmentations aims
to introduce more speaker variability by the removal
of random croppings from the audio signal to enlarge
the dataset, the adjustment of the loudness volume of
some segments (i.e., multiplication by a constant in
the range [0.5, 2]), the modification of the pitch with
a factor in the range [−10, 10], the modification of
the speed (i.e., time scaling with a factor in the range
[0.5, 2]), and Vocal Tract Length Perturbation (VTLP)
with a factor in the range [0.9, 1.1].
4.2 CCNU Mobile
The Central China Normal University (CCNU) Mo-
bile dataset (Zeng et al., 2020) consists of speech
recordings from the TIMIT dataset recorded by 45
Mobile Phone Identification from Recorded Speech Signals Using Non-Speech Segments and Universal Background Model Adaptation
795

mobile phones belonging to 9 brands. Multiple de-
vices of the same mobile phone model are included
in the dataset. The recordings of each phone consist
of 642 audio samples per mobile phone with a dura-
tion of about 8 seconds sampled at 32kHz. Unfortu-
nately, the entire dataset is not publicly available. We
were able to use a small subset of the dataset, which
consists of 5 8-second-long audio samples of female
speakers per mobile phone. This data scarcity will al-
low us to test our methods under realistic audio foren-
sic conditions with limited data.
5 EXPERIMENTAL RESULTS
5.1 Results on MOBIPHONE
Experiments were conducted on the MOBIPHONE
dataset using the traditional, non-speech, and UBM
approaches. Two tasks were considered, namely
brand identification and model identification. We
used a 50%/50% train/test split with half male and
half female speakers in both sets.
23 MFCCs were extracted using 20 ms Hamming
windows with 10 ms overlap. The MFCCs for the
non-speech approach were extracted from non-speech
frames obtained via voice activity detection. The
MFCCs for the traditional and UBM approaches are
referred to as audio MFCCs, while the MFCCs for
the non-speech approach are referred to as non-speech
MFCCs.
Model- and brand-dependent GMMs were trained
via the EM algorithm, using a k-means++ initializa-
tion for both the audio and non-speech MFCC vec-
tors. We trained GMMs with the number of com-
ponents in the range 1, 2, 4, 8, 16, 32, and 64.
These GMMs were used in the ML classification. For
the SVM classification, utterance-dependent GMMs
were trained using audio and non-speech MFCC vec-
tors. The components of these GMMs were in the
range 1, 2, and 4 because the classification accuracy
was drastically reduced when more than 4 compo-
nents were employed in the extracted GSVs due to
the large number of features.
Finally, utterance-dependent GMMs were also
trained using UBM MAP adaptation applied to the au-
dio MFCCs. The UBMs were trained on the TIMIT
dataset. In the case of the UBM-adapted GSVs, it
was seen that SVMs with UBMs up to 16 components
yielded good results. As is most common, only the
means were adapted.
The results of the initial experiments on the MO-
BIPHONE dataset are listed in Tables 1-4. The
first column shows the number of components of the
GMMs. The remaining columns display the accu-
racies of the traditional approach (Audio), the non-
speech approach (Non-speech), and the UBM ap-
proach (UBM) on the corresponding classification
method.
Table 1: ML classification accuracy for brand identification
on the MOBIPHONE dataset using audio and non-speech
MFCCs.
# Components Audio Non-speech
1 66.6667% 79.3651%
2 73.4127% 86.1111%
4 88.4921% 96.8254%
8 91.6667% 99.2063%
16 88.0952% 99.6032%
32 93.254% 100%
64 98.0159% 99.6032%
Table 2: ML classification accuracy for model identification
on the MOBIPHONE dataset using audio and non-speech
MFCCs.
# Components Audio Non-speech
1 96.0317% 98.8095%
2 98.8095% 99.6032%
4 98.4127% 100%
8 99.2063% 100%
16 98.4127% 100%
32 98.8095% 100%
64 98.4127% 100%
A first remark on the results in Tables 1-4 is that
the brand identification problem seems harder than
the model identification problem. In practice, the
brand can be easily deduced from the model. Since
model identification performs better, brand identifica-
tion is of no practical use. However, by examining it
we see the accuracy increase offered by non-speech
MFCCs and UBM adaptation. The non-speech and
the UBM approaches perform better than the tradi-
tional approach when performing ML or SVM clas-
sification. In the non-speech approach, we also note
that the GSVs with the covariance don’t always yield
an accuracy increase compared to the mean GSVs.
This can also be verified in the following experiments.
The same classification procedure was followed
on the augmented versions of the MOBIPHONE
dataset. The aim is to demonstrate how the dif-
ferent approaches perform under different types of
noise. The 504 samples of the baseline MOBI-
PHONE dataset were expanded by the 504 samples
of one or more augmented versions.
The best accuracy on the first set of augmenta-
tions, which simulate different recording conditions,
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
796

Table 3: SVM classification accuracy for brand identification on the MOBIPHONE dataset using GSVs extracted from audio
MFCCs, non-speech MFCCs, and UBM adaptation.
Audio Non-speech UBM
# Components Means Means+Cov Means Means+Cov Means
1 93.6508% 96.4286% 97.619% 97.619% 98.0159%
2 93.6508% 96.0317% 97.619% 97.619% 97.2222%
4 82.5397% 85.3175% 92.8571% 92.8571% 98.8095%
8 - - - - 98.8095%
16 - - - - 96.0317%
Table 4: SVM classification accuracy for model identification on the MOBIPHONE dataset using GSVs extracted from audio
MFCCs, non-speech MFCCs, and UBM adaptation.
Audio Non-speech UBM
# Components Means Means+Cov Means Means+Cov Means
1 94.0476% 96.4286% 99.2063% 98.8095% 94.0476%
2 94.0476% 96.4286% 97.619% 97.2222% 98.0159%
4 92.8571% 94.4444% 96.4286% 95.2381% 98.8095%
8 - - - - 98.8095%
16 - - - - 94.4444%
Table 5: The best classification accuracy on MOBIPHONE for brand identification on the first set of augmentations.
Audio Non-speech UBM
ML SVM ML SVM SVM
Baseline 98.0159%
64
96.4286%
1/MC
100%
32
97.619%
1/M
98.8095%
4/M
Gaussian 85.7143%
64
89.2857%
1/M
81.3492%
64
92.2619%
1/MC
89.2857%
1/M
Background 92.0635%
64
92.8571%
2/M
86.5079%
64
91.6667%
1/M
98.4127%
4/M
Reverberation 96.8254%
32
95.0397%
1/MC
99.0079%
64
97.2222%
1/MC
98.4127%
8/M
All 88.8889%
64
93.0556%
1/MC
83.3333%
64
90.4762%
1/M
95.2381%
2/M
Table 6: The best classification accuracy on MOBIPHONE for model identification on the first set of augmentations.
Audio Non-speech UBM
ML SVM ML SVM SVM
Baseline 99.2063%
8
96.4286%
1/MC
100%
4
99.2063%
1/M
98.8095%
4/M
Gaussian 89.881%
16
90.6746%
1/MC
78.9683%
64
93.8492%
1/M
92.8571%
2/M
Background 93.0556%
16
95.4365%
1/MC
83.3333%
64
94.0476%
1/MC
97.4206%
2/M
Reverberation 98.6111%
8
96.2302%
2/MC
99.6032%
32
98.4127%
1/M
98.2143%
4/M
All 91.5675%
64
92.2619%
1/M
80.2579%
64
91.8651%
1/MC
96.2302%
2/M
is summarized in Tables 5-6 for brand and model
identification, respectively. The best hyperparameters
are indicated, namely the number of Gaussian com-
ponents and the type of GSVs (means/means and co-
variance). The best accuracy listed in Tables 1-4 is
quoted in the row Baseline of Tables 5-6.
For brand and model identification, non-speech
ML classification performs worse than the traditional
ML classification in all cases except for the reverber-
ation augmentation. However, non-speech SVM clas-
sification performs better than the traditional SVM
classification in the Gaussian noise and reverberation
augmentations. In all cases, the UBM-adapted SVM
classification achieves higher accuracy than the tradi-
tional SVM classification. It also yields the top ac-
curacy when all the augmentations are used together.
Overall, the UBM approach performs better on the
first set of augmentations. This is attributed to the
probabilistic alignment of the MFCC vectors to the
UBM, which can mitigate the effects of the noise on
the features.
The best classification accuracy on the second set
of augmentations is quoted in Tables 7-8 for brand
and model identification, respectively. The non-
speech ML classification of brand and model achieves
the highest accuracy for all augmentations but the
VTLP augmentation in the model identification task.
Speaker variability is eliminated because the non-
speech features are not affected much by the speaker-
dependent noise amplified with this set of augmen-
tations. When SVM classification is employed, the
UBM and non-speech approaches achieve better ac-
Mobile Phone Identification from Recorded Speech Signals Using Non-Speech Segments and Universal Background Model Adaptation
797

Table 7: The best classification accuracy on MOBIPHONE for brand identification on the second set of augmentations.
Audio Non-speech UBM
ML SVM ML SVM SVM
Baseline 98.0159%
64
96.4286%
1/MC
100%
32
97.619%
1/M
98.8095%
4/M
Crop 98.2143%
64
96.8254%
1/MC
99.8016%
64
98.6111%
1/M
99.0079%
16/M
Loudness 97.4206%
64
96.0317%
1/MC
99.0079%
8
99.2063%
1/M
97.8175%
2/M
Pitch 92.0635%
64
86.3095%
1/MC
97.4206%
64
94.8413%
2/MC
94.4444%
16/M
Speed 98.2143%
64
97.0238%
2/MC
99.6032%
32
99.2063%
1/MC
98.4127%
16/M
VTLP 96.4286%
64
98.0159%
1/MC
98.4127%
32
97.619%
1/MC
97.619%
8/M
All 95.6349%
64
95.172%
1/MC
98.0159%
64
96.627%
1/MC
96.9577%
8/M
Table 8: The best classification accuracy on MOBIPHONE for model identification on the second set of augmentations.
Audio Non-speech UBM
ML SVM ML SVM SVM
Baseline 99.2063%
8
96.4286%
1/MC
100%
4
99.2063%
1/M
98.8095%
4/M
Crop 99.4048%
32
96.2302%
2/MC
100%
4
99.0079%
1/MC
99.6032%
4/M
Loudness 98.8095%
8
97.4206%
1/MC
99.2063%
8
99.0079%
1/M
99.0079%
2/M
Pitch 94.6429%
64
86.9048%
1/MC
98.4127%
32
97.0238%
1/MC
97.4206%
4/M
Speed 98.8095%
8
96.627%
1/MC
100%
4
99.4048%
1/MC
98.0159%
8/M
VTLP 99.4048%
64
98.0159%
1/MC
98.8095%
4
99.2063%
1/MC
98.8095%
4/M
All 97.2884%
64
93.1217%
1/M
98.8095%
32
96.6931%
1/M
98.8095%
8/M
Table 9: The best classification accuracy on MOBIPHONE for brand identification on both sets of augmentations combined.
Audio Non-speech UBM
ML SVM ML SVM SVM
Baseline 98.0159%
64
96.4286%
1/MC
100%
32
97.619%
1/M
98.8095%
4/M
All augmentations 91.4021%
64
91.9312%
1/MC
89.903%
64
93.9594%
1/M
93.6067%
2/M
Table 10: The best classification accuracy on MOBIPHONE for model identification on both sets of augmentations combined.
Audio Non-speech UBM
ML SVM ML SVM SVM
Baseline 99.2063%
8
96.4286%
1/MC
100%
4
99.2063%
1/M
98.8095%
4/M
All augmentations 93.739%
64
92.0635%
2/MC
89.9471%
64
94.0035%
1/M
93.2099%
2/M
Table 11: Mean and standard deviation of the classification
accuracy for brand identification on the neural network us-
ing Zeng’s UBM.
Zeng’s UBM
MFCC branch 41.7777% (5.1225%)
GSV branch 91.4074% (0.9999%)
Fusion branch 70.7407% (7.7%)
Table 12: Mean and standard deviation of the classification
accuracy for model identification on the neural network us-
ing Zeng’s UBM.
Zeng’s UBM
MFCC branch 23.9555% (6.6566%)
GSV branch 96.2222% (1.54%)
Fusion branch 73.8888% (10.2003%)
curacy than the traditional approach.
When both sets of augmentations are used to-
gether, the brand and model identification accuracy
is summarized in Tables 9-10. In both tasks, non-
speech SVMs yield the best accuracy. Non-speech
ML classification performs worse than the traditional
ML classification. UBM-adapted SVM classification
also performs better than the traditional SVM classi-
fication.
5.2 Results on CCNU Mobile
The CCNU Mobile dataset was used to compare the
identification methods further. In addition to the ML
and SVM classification methods, the neural architec-
ture designed specifically for mobile phone identifi-
cation in (Zeng et al., 2023) was included.
The entirety of the CCNU Mobile dataset is not
available, and we could use only a small sample of
it containing 5 audio samples from female speakers
for each of the 45 mobile phones. However, the lim-
ited data helps us test the quality of the extracted data
since the feature vectors are so few the classifiers have
to rely on the data’s quality and not quantity to yield
better results.
The first two speakers were chosen for the train
set, while the remaining 3 speakers were selected for
the test set. To be consistent with the features used
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
798

Table 13: Mean and standard deviation of the classification accuracy for brand identification using the three approaches with
ML, SVMs, and the neural network proposed by (Zeng et al., 2023).
Audio Non-speech UBM
ML-64 86.5899% (1.7982%) 99.2593% (0%) -
SVM-1M 86.0741% (4.2881%) 86.8148 % (5.8514%) -
SVM-1MC 68.6667% (10.0482%) 83.4074% (4.7263%) -
SVM-4M - - 86.5926% (1.3726%)
MFCC branch 45.3333% (6.5309%) 60.3703% (6.6369%) 44.5185% (3.7357%)
GSV branch 61.0370% (3.7803%) 83.3333% (2.5241%) 94.7407% (1.3725%)
Fusion branch 60.5926% (4.9481%) 81.6296% (3.9009%) 70.0740% (7.3761%)
Table 14: Mean and standard deviation of the classification accuracy for model identification using the three approaches with
ML, SVMs, and the neural network proposed by (Zeng et al., 2023).
Audio Non-speech UBM
ML-64 98.5185% (0%) 98.5185% (0%) -
SVM-1M 90.8148% (5.7292%) 96.5185% (3.5447%) -
SVM-1MC 81.9259% (6.3356%) 90.7407 % (1.0030%) -
SVM-4M - - 93.3333% (0.7808%)
MFCC branch 22.3703% (5.2471%) 43.7037% (7.7454%) 21.7037% (5.3307%)
GSV branch 56.7407% (4.3361%) 89.8518% (2.9845%) 98.2222% (1.1152%)
Fusion branch 52.1481% (2.0717%) 86.4444% (4.3199%) 63.5555% (13.1018%)
in (Zeng et al., 2023), we also used 39-dimensional
MFCC vectors. The features were extracted using 16
ms Hamming windows with an 8 ms overlap.
Because the audio samples were only 8 seconds
long, the voice activity detection used in the non-
speech approach did not produce enough frames for
the GMM training. So, for the non-speech approach,
residual noise signals were obtained by subtracting
an estimated clean speech signal from the original
speech signal.
To enable a fair comparison between the accu-
racy of our approaches to those disclosed in (Zeng
et al., 2023), we emulated the results that would have
been obtained on the small sample of the CCNU Mo-
bile dataset by using the UBM trained on the TIMIT
dataset of the original paper. In every experiment, the
neural network training is done for 100 epochs using
a batch size of 16. The learning rate is 0.001 and de-
creases by 1/10 every 20 epochs. The Adaptive Mo-
ment Estimator (Adam) optimizer is used. Brand and
model identification accuracy using Zeng’s UBM are
shown in Tables 11-12. The GMM hyperparameters
were fixed for this experiment, so every classification
was repeated 10 times and the mean and standard de-
viation were reported.
The MFCC branch of the network is the module
that works with the MFCCs only, the GSV branch is
the module that works with the GSVs only, and the
fusion branch is the module that fuses the results of
the previous two branches to obtain a final classifica-
tion. Implementation details can be found in (Zeng
et al., 2023). It is seen that without enough data, the
GSV branch performs better than the MFCC branch
and the fusion branch in both tasks.
The results of the traditional, the non-speech, and
the UBM approach, when ML, SVM, and neural clas-
sification are applied, are summarized in Tables 13-
14. The UBM here is also trained on the TIMIT
dataset. For the ML classification, 64 component
GMMs are used. For the GSVs fed to the SVMs, 1
component is used, while for the SVMs applied to the
UBM-adapted GSVs, 4 components are used.
In both tasks, the non-speech MFCCs improve the
accuracy of all classification approaches compared to
the traditional approach. This is also true when the
UBM is used. Examining the neural network clas-
sification a little closer (i.e., the last 3 table rows in
Tables 13 and 14), we see that the GSV branch us-
ing the UBM-adapted GSVs gives the best mean ac-
curacy for both brand and model identification when
the neural classification is employed. Zeng’s UBM
(Tables 11-12) yields similar results to our UBM, as
expected, except the fusion branch for model iden-
tification, where Zeng’s UBM attains a higher mean
accuracy but a similar large standard deviation.
6 CONCLUSIONS
This paper has disclosed empirical evidence demon-
strating that either the non-speech or UBM ap-
proaches increase mobile phone identification classi-
fication accuracy, even under data scarcity and noisy
recordings. The non-speech approach performs better
when a lot of speaker variability is observed since the
features extracted from the non-speech segments of
the signal are not corrupted by speaker-related noise.
The UBM approach performs better in the presence of
recording noise because of the probabilistic “match-
ing” of the extracted features with the UBM’s com-
ponents, which is robust to noisy features.
Considering future work, more experiments
should be conducted on other datasets, especially the
entire CCNU Mobile dataset if it becomes publicly
Mobile Phone Identification from Recorded Speech Signals Using Non-Speech Segments and Universal Background Model Adaptation
799

available, to check whether our results on the small
sample are similar to the whole dataset. In the non-
speech approach, instead of the MFCCs, which are
designed based on the logarithmic scale of the human
auditory system, other features providing the same
resolution in all frequency ranges might also yield
better results since the features are extracted from
noise-like signals and not speech.
ACKNOWLEDGEMENTS
This research was supported by the Hellenic Foun-
dation for Research and Innovation (H.F.R.I.) under
the “2nd Call for H.F.R.I Research Projects to support
Faculty Members & Researchers” (Project Number:
3888).
REFERENCES
Alsulaiman, M., Muhammad, G., Bencherif, M. A., Mah-
mood, A., and Ali, Z. (2013). Ksu rich arabic speech
database. Information (Japan), 16(6 B):4231–4253.
Baldini, G. and Amerini, I. (2019). Smartphones identi-
fication through the built-in microphones with con-
volutional neural network. IEEE Access, 7:158685–
158696.
Baldini, G. and Amerini, I. (2022). Microphone identifica-
tion based on spectral entropy with convolutional neu-
ral network. In Proc. 2022 IEEE International Work-
shop on Information Forensics and Security (WIFS),
pages 1–6. IEEE.
Berdich, A., Groza, B., Levy, E., Shabtai, A., Elovici,
Y., and Mayrhofer, R. (2022). Fingerprinting smart-
phones based on microphone characteristics from
environment affected recordings. IEEE Access,
10:122399–122413.
Berdich, A., Groza, B., and Mayrhofer, R. (2023). A survey
on fingerprinting technologies for smartphones based
on embedded transducers. IEEE Internet of Things
Journal, 10(16):14646–14670.
Davis, S. and Mermelstein, P. (1980). Comparison of para-
metric representations for monosyllabic word recog-
nition in continuously spoken sentences. IEEE Trans-
actions on Acoustics, Speech, and Signal Processing,
28(4):357–366.
Dempster, A., Laird, N., and Rubin, D. (1977). Maxi-
mum likelihood from incomplete data via the em algo-
rithm. Journal of the Royal Statistical Society: Series
B (methodological), 39(1):1–22.
Garcia-Romero, D. and Espy-Wilson, C. Y. (2010). Au-
tomatic acquisition device identification from speech
recordings. In Proc. 2010 IEEE International Con-
ference on Acoustics, Speech, and Signal Processing,
pages 1806–1809. IEEE.
Garofolo, J. S., Lamel, L. F., Fisher, W. M., Fiscus, J. G.,
and Pallett, D. S. (1988). Getting started with the
darpa timit cd-rom: An acoustic phonetic continuous
speech database. National Institute of Standards and
Technology (NIST), Gaithersburgh, MD, 107:16.
Giganti, A., Cuccovillo, L., Bestagini, P., Aichroth, P., and
Tubaro, S. (2022). Speaker-independent microphone
identification in noisy conditions. In Proc. 2022 30th
European Signal Processing Conference (EUSIPCO),
pages 1047–1051. IEEE.
Hanilci, C., Ertas, F., Ertas, T., and Eskidere,
¨
O. (2011).
Recognition of brand and models of cell-phones from
recorded speech signals. IEEE Transactions on Infor-
mation Forensics and Security, 7(2):625–634.
Hanilc¸i, C. and Kinnunen, T. (2014). Source cell-phone
recognition from recorded speech using non-speech
segments. Digital Signal Processing, 35:75–85.
Khan, M. K., Zakariah, M., Malik, H., and Choo, K.-K. R.
(2018). A novel audio forensic data-set for digital
multimedia forensics. Australian Journal of Forensic
Sciences, 50(5):525–542.
Kotropoulos, C. (2014). Source phone identification using
sketches of features. IET Biometrics, 3(2):75–83.
Kotropoulos, C. and Samaras, S. (2014). Mobile phone
identification using recorded speech signals. In Proc.
2014 19th International Conference on Digital Signal
Processing, pages 586–591. IEEE.
Leonzio, D. U., Cuccovillo, L., Bestagini, P., Marcon, M.,
Aichroth, P., and Tubaro, S. (2023). Audio splicing
detection and localization based on acquisition device
traces. IEEE Transactions on Information Forensics
and Security, 18:4157–4172.
Maher, R. (2009). Audio forensic examination. IEEE Signal
Processing Magazine, 26(2):84–94.
Martin, R. (2001). Noise power spectral density estimation
based on optimal smoothing and minimum statistics.
IEEE Transactions on Speech and Audio Processing,
9(5):504–512.
Qamhan, M., Alotaibi, Y. A., and Selouani, S. A. (2023).
Transformer for authenticating the source microphone
in digital audio forensics. Forensic Science Interna-
tional: Digital Investigation, 45:301539.
Reynolds, D., Quatieri, T., and Dunn, R. (2000). Speaker
verification using adapted gaussian mixture models.
Digital Signal Processing, 10(1-3):19–41.
Reynolds, D. A. (1997). HTIMIT and LLHDB: speech cor-
pora for the study of handset transducer effects. In
Proc. 1997 IEEE International Conference on Acous-
tics, Speech, and Signal Processing, volume 2, pages
1535–1538. IEEE.
Sohn, J., Kim, N. S., and Sung, W. (1999). A statistical
model-based voice activity detection. IEEE Signal
Processing Letters, 6(1):1–3.
Zeng, C., Feng, S., Zhu, D., and Wang, Z. (2023). Source
acquisition device identification from recorded audio
based on spatiotemporal representation learning with
multi-attention mechanisms. Entropy, 25(4):626.
Zeng, C., Zhu, D., Wang, Z., Wang, Z., Zhao, N., and
He, L. (2020). An end-to-end deep source record-
ing device identification system for web media foren-
sics. International Journal of Web Information Sys-
tems, 16(4):413–425.
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
800