Speaker Verification Enhancement via Speaking Rate Dynamics in
Persian Speechprints
Nina Hosseini-Kivanani
1 a
, Homa Asadi
2 b
and Christoph Schommer
1 c
1
Department of Computer Science, University of Luxembourg, Esch-sur-Alzette, Luxembourg
2
Faculty of Foreign Languages, University of Isfahan, Isfahan, Iran
Keywords:
Speaker Verification, Mel-Frequency Cepstral Coefficients (MFCCs), Vowel Formants, Deep Learning,
Persian Language.
Abstract:
This paper investigates the impact of speaking rate variation on speaker verification using a hybrid feature
approach that combines Mel-Frequency Cepstral Coefficients (MFCCs), their dynamic derivatives (delta and
delta-delta), and vowel formants. To enhance system robustness, we also applied data augmentation tech-
niques such as time-stretching, pitch-shifting, and noise addition. The dataset comprises recordings of Persian
speakers at three distinct speaking rates: slow, normal, and fast. Our results show that the combined model
integrating MFCCs, delta-delta features, and formant frequencies significantly outperforms individual fea-
ture sets, achieving an accuracy of 75% with augmentation, compared to 70% without augmentation. This
highlights the benefit of leveraging both spectral and temporal features for speaker verification under varying
speaking conditions. Furthermore, data augmentation improved the generalization of all models, particularly
for the combined feature set, where precision, recall, and F1-score metrics showed substantial gains. These
findings underscore the importance of feature fusion and augmentation in developing robust speaker veri-
fication systems. Our study contributes to advancing speaker identification methodologies, particularly in
real-world applications where variability in speaking rate and environmental conditions presents a challenge.
1 INTRODUCTION
Speech production is a highly complex phenomenon
in which dynamic articulatory gestures drive the
movements of speech organs to achieve specific tar-
gets within the vocal tract geometry (Tilsen, 2014).
These articulatory movements shape the acoustic fea-
tures of speech, which carry rich information, en-
abling listeners to comprehend both the linguistic
content (what is said) and the speaker-specific de-
tails (who said it). Identifying speakers based on the
characteristics of their voices is a fundamental goal in
forensic phonetics and automatic speaker recognition
systems (Rose, 2002; Nolan, 1987).
One of the key challenges in speaker identifica-
tion is the high variability in acoustic characteristics
across speakers, compared to the relatively low vari-
ability within a single speaker (Gold et al., 2013;
McDougall, 2006). In forensic speaker comparison
a
https://orcid.org/0000-0002-0821-9125
b
https://orcid.org/0000-0003-1655-1336
c
https://orcid.org/0000-0002-0308-7637
(FSC), addressing this variability is crucial when de-
termining whether a known voice sample matches
an unknown (disputed) sample (Rose, 2002; Nolan,
1987).
Identifying robust speaker-specific parameters re-
mains challenging due to intertwined factors, such
as linguistic effects, prosody, and channel condi-
tions, which influence system accuracy (Rose, 2002).
Among these factors, speaking rate variability plays a
particularly important role, as speakers naturally ad-
just their rate based on communicative context, emo-
tional state, or physiological conditions (Gay et al.,
1974; Imaizumi and Kiritani, 1989). Such variations
affect both articulatory movements and acoustic prop-
erties, posing challenges to speaker recognition sys-
tems (Shahrebabaki et al., 2018; Zeng et al., 2015).
From a computational standpoint, automatic
speaker identification (ASI) systems must account for
changes in speaking rate (Reynolds et al., 2000). Tra-
ditional ASI systems use MFCCs as their primary fea-
ture set due to their effectiveness in capturing speaker-
specific spectral information(Davis and Mermelstein,
1980). However, these systems often degrade when
Hosseini-Kivanani, N., Asadi, H. and Schommer, C.
Speaker Verification Enhancement via Speaking Rate Dynamics in Persian Speechprints.
DOI: 10.5220/0013189100003905
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 665-672
ISBN: 978-989-758-730-6; ISSN: 2184-4313
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
665

faced with speaking rate variations, as spectral char-
acteristics shift with changes in articulatory dynam-
ics(Zeng et al., 2015). Recent advances in deep learn-
ing (DL) have addressed this by developing speaker-
invariant representations that are more robust to such
variations (Hinton et al., 2012; Xie et al., 2019).
Despite these advancements, vowel formants re-
main valuable in forensic applications due to their in-
terpretability and close relationship to the physiolog-
ical aspects of speech production (Gold et al., 2013).
Given that both vowel formants and MFCCs are dif-
ferentially affected by speaking rate, assessing their
relative robustness is critical for improving speaker
identification systems.
In this study, we evaluate the impact of speaking
rate variability on speaker identification performance
by comparing the robustness of vowel formants and
MFCCs. Specifically, we aim to:
1. examine how speaking rate influences vowel for-
mant frequencies and MFCC features;
2. assess the effectiveness of these features in
speaker identification under varying speaking
rates;
3. and determine whether combining formant and
MFCC features improves accuracy and resilience
in speaker verification systems.
By addressing these objectives, this study pro-
vides insights into selecting acoustic features most
effective for speaker identification under conditions
of variable speaking rates. The remainder of this pa-
per is organized as follows: Section 2 reviews related
work from both phonetic and computational perspec-
tives. Section 3 outlines our experimental methodol-
ogy. Section 4 presents our findings, while Section 5
discusses their implications. Finally, Section 6 con-
cludes the paper and suggests directions for future re-
search.
2 RELATED WORK
Forensic speaker comparison (FSC) and automatic
speaker recognition systems rely on acoustic features,
such as vowel formants and Mel-Frequency Cep-
stral Coefficients (MFCCs), to differentiate speak-
ers based on their unique vocal characteristics (Rose,
2002; Nolan, 1987). These features effectively cap-
ture speaker-specific and linguistically relevant as-
pects of speech, making them central to forensic and
automatic speaker verification tasks.
Speaking rate is a critical variable that affects the
articulation and acoustic properties of speech. Re-
search has shown that varying speaking rates influ-
ence the kinematics of speech production. For in-
stance, Gay (Gay et al., 1974) demonstrated that in-
creased speaking rate is associated with heightened
muscle activity, such as more pronounced lip clo-
sure and greater bilabial consonant openings. Sim-
ilarly, Tuller and Kelso (Tuller et al., 1982) found
that faster-speaking rates result in shorter muscle ac-
tivity durations, while slower rates lead to longer ar-
ticulatory movement times. These findings under-
score how speaking rate directly impacts speech ar-
ticulation, making a vital factor in speaker verifica-
tion tasks. These articulatory changes, influenced
by speaking rate, are further reflected in the vari-
ability of specific speech gestures. For instance,
Shaiman et al. (Shaiman et al., 1997) observed that
lip gestures exhibit variability across different speak-
ing rates, complicating the consistency of speaker-
specific features as articulation velocity changes.
From an acoustic perspective, speaking rate sig-
nificantly alters the spectral characteristics and for-
mant trajectories of speech. Imaizumi and Kiri-
tani (Imaizumi and Kiritani, 1989) observed that rapid
speech can lead to vowel reduction, especially in the
second formant frequency (F2). Additionally, Weis-
mer and Berry (Weismer and Berry, 2003) showed
that speakers modify formant movement trajectories
based on their speaking rate, with F2 being particu-
larly affected by these changes. This suggests that
variations in speaking rate not only affect articulatory
gestures but also modify key acoustic features critical
for speaker recognition.
Furthermore, research by Mefferd and
Green (Mefferd and Green, 2010) demonstrated
that formant transitions become sharper at higher
speaking rates. Their studies also noted that vowel
formant distances exhibit greater specificity in slow
speech compared to fast speech. Agwuele et al. (Ag-
wuele et al., 2009) found a reduction in vowel space
with faster speaking rates, indicating that articulation
and vowel acoustics are closely tied to the speed
of speech. Together, these studies highlight the
variability in formant frequency patterns and their
dependency on speaking rate, a critical challenge for
speaker identification systems.
Recent advancements in speaker verification have
explored the use of learnable MFCCs to improve ro-
bustness to variable speaking rates. For instance, Liu
et al. (Liu et al., 2021) introduced adaptive MFCC
front-end architectures that adjust to data, making
these features more resilient to changing speech con-
ditions, including speaking rate variability. This adap-
tive approach has shown significant improvements
in speaker verification performance, particularly in
large-scale datasets like VoxCeleb1 and SITW. How-
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
666

ever, while learnable MFCCs offer improvements,
they may not fully capture the dynamic properties of
speech under all conditions.
2.1 Speaker Recognition and
Computational Approaches
In traditional automatic speaker recognition, MFCCs
are widely used for their robustness in capturing
speaker-specific spectral properties (Davis and Mer-
melstein, 1980; Reynolds et al., 2000). However, they
are susceptible to degradation under varying speak-
ing rates, as spectral characteristics shift due to ar-
ticulatory dynamics (Zeng et al., 2015). Zeng and
Sheng (Zeng et al., 2015) demonstrated that these
changes directly affect the reliability of MFCC-based
systems. To address these limitations, recent ap-
proaches have incorporated machine learning (ML)
and deep learning (DL) techniques, such as convo-
lutional neural networks (CNNs) and utterance-level
aggregation methods (Xie et al., 2019), which learn
hierarchical representations that improve the robust-
ness of speaker verification systems under speaking
rate variability (Hinton et al., 2012).
Recent work has also explored hybrid models
that integrate the features of MFCCs with DL ar-
chitectures. For instance, combining MFCCs with
CNNs captures both local and global speech pat-
terns, enhancing robustness to noise and rate variabil-
ity. Furthermore, bi-directional long short-term mem-
ory (Bi-LSTM) networks improve speaker verifica-
tion by modeling long-range temporal dependencies,
which are essential for capturing dynamic speech
variations (Anupama et al., 2022).
Hybrid optimization strategies have fur-
ther advanced these models. Chakravarty &
Dua (Chakravarty and Dua, 2023) combined MFCCs
and Gammatone Cepstral Coefficients (GTCCs)
with data augmentation methods, such as Synthetic
Minority Over-Sampling Technique (SMOTE), to
improve performance. By leveraging a hybrid LSTM
backend, their approach enhanced model accuracy
and resilience under noisy conditions, demonstrating
the value of combining feature sets and optimization
algorithms in varying conditions.
In forensic applications, interpretability remains
critical, making vowel formants valuable despite their
susceptibility to speaking rate variability. Formants
offer insights into the physiological aspects of speech
production, which are useful for distinguishing speak-
ers (Asadi et al., 2018; McDougall, 2006). How-
ever, their limitations as standalone features neces-
sitate combining formants with robust features like
MFCCs. Jahangir et al. (Jahangir et al., 2020)
demonstrated that integrating traditional acoustic fea-
tures with DL-derived representations significantly
enhances speaker identification under variable speak-
ing rates.
Some studies have also shown that the fusion
of acoustic features improves performance. Bahari
et al.(Bahari and Van Hamme, 2011) demonstrated
that combining formant frequencies and MFCCs cap-
tures both articulatory and spectral information, lead-
ing to better recognition rates under challenging con-
ditions. Advanced modeling techniques, such as i-
vectors and x-vectors, have further demonstrated ro-
bustness across varying speaking conditions(Dehak
et al., 2010). These findings emphasize the benefits
of combining diverse feature sets for robust speaker
verification systems.
2.2 Research Gap
While progress has been made in understanding how
speaking rate affects speech production and acous-
tic features, few studies comprehensively compare the
robustness of formants and MFCCs in speaker verifi-
cation under varying rates. Further research is needed
to evaluate whether combining these features with
deep learning techniques can improve resilience to
speaking rate variability, especially in forensic and
real-world applications. In this study, we address
this gap by systematically examining the impact of
speaking rate on formant frequencies and MFCCs in
speaker verification tasks. Additionally, we explore
the benefits of integrating these features into a unified
framework to improve accuracy and robustness under
variable speaking conditions.
3 EXPERIMENTAL SET-UP
3.1 Participants and Task
Eighteen male Persian speakers (Tehrani variety; age
range: 25–36 years; M = 31.3, SD = 3.7) were
recorded. None of the speakers reported any hearing
or speech impairments. All participants were students
pursuing a master’s or PhD degree in various research
areas. This corpus was collected following the pro-
cedure used in the collection of the BonnTempo cor-
pus in German. Speakers were instructed to read The
North Wind and the Sun in Persian at three different
speaking rates (slow, normal, and fast). Before each
recording session, participants were asked to read the
text several times to familiarize themselves with the
passage. First, speakers were instructed to read the
passage at their normal pace. The speakers were then
Speaker Verification Enhancement via Speaking Rate Dynamics in Persian Speechprints
667

asked to slow their pace as much as they could and
then to read the text as fast as possible. This resulted
in strong syllable rate variability across the three dif-
ferent reading passages. All recording sessions took
place in a soundproof booth with a sampling rate of
44.1 kHz and 16-bit quantization.
3.2 Feature Extraction
Speech recordings were labeled and segmented
based on the onset and offset information using
Praat (Boersma and Weenink, 2021) version 6.2.22.
A free plugin for Praat with automated scripts for
voice processing, Praat Vocal Toolkit (Corretge,
2022), was used to extract and concatenate all vow-
els from each recording per speaker. Formant values
were extracted at 5-ms intervals using the LPC-based
Burg algorithm in Praat. A long-term analysis method
was adopted, as it has proven effective in represent-
ing speaker individuality (Asadi et al., 2018; Gold
et al., 2013). This approach calculates the average for-
mant values over a long stretch of a speaker’s speech
recording (Gold et al., 2013; Rose, 2002; Nolan,
1987).
The key features were extracted using MFCCs
in Python (Version 3.11.5) with the librosa li-
brary (McFee et al., 2015). Thirteen main coefficients
were calculated and averaged per audio file to sim-
plify the representation while preserving key sound
characteristics.
We developed a multi-class speech analysis model
that extensively uses MFCCs, delta-MFCCs, and
delta-delta-MFCCs to capture a broad spectrum of
acoustic features from speech. This strategy enriches
the model with spectral and temporal information,
improving its ability to distinguish between varied
speaking speeds and speaker characteristics. The
methodology centers on extracting a rich set of fea-
tures from audio recordings: the base MFCCs pro-
vide spectral information, delta-MFCCs capture the
rate of change in these spectral features, and delta-
delta-MFCCs further detail the acceleration of these
changes. This layered approach to feature extraction
ensures a deep representation of the audio’s charac-
teristics. Each feature dimension is normalized to en-
sure consistency in scale across the dataset, facilitat-
ing more effective model training.
Adding attention mechanisms, such as self-
attention or Transformer layers, can help the model
focus on the most relevant parts of the speech signal
for speaker verification, making it more effective in
handling variations in speaking speed. We then trans-
formed these labels into a one-hot encoded format
suitable for classification tasks. The dataset was split
into training and testing sets, with 20% reserved for
testing to evaluate the model’s performance. Our neu-
ral network model architecture included Long Short-
Term Memory (LSTM) layers to process the sequen-
tial nature of the audio data, followed by a custom At-
tentionLayer designed to weigh the importance of dif-
ferent parts of the audio signal, enhancing the model’s
focus on relevant features for speaker verification.
The model also incorporated dropout layers to prevent
overfitting and used a softmax activation function in
the output layer for classification. The model was
compiled using the Adam optimizer and categorical
cross-entropy loss. Early stopping was implemented
to terminate training when validation loss ceased to
improve, thereby preventing overfitting
We used Group K-Fold Cross-Validation, a variant
of K-Fold, to divide the dataset into five folds while
ensuring that all recordings from a single speaker
were placed either in the training or testing set, not
both. This approach prevents data leakage and en-
sures a fair evaluation by maintaining class propor-
tions across folds, allowing the model to generalize
effectively to unseen speakers.
3.3 Speech Data Augmentation
To enhance dataset diversity and improve model gen-
eralization, five audio augmentation techniques were
applied using the librosa library. Augmentation in-
troduces variability in training data, helping models
better handle real-world challenges such as varying
speaking rates, background noise, and recording con-
ditions (Lounnas et al., 2022).
Time-Stretching: We adjusted the speed of the au-
dio using factors of 0.9 (slower) and 1.1 (faster). This
manipulation simulates natural variations in speaking
rate, allowing the model to adapt to speakers with dif-
ferent articulation speeds (Ko et al., 2015).
Pitch Shifting: The pitch of the audio was shifted
by ±2 semitones to mimic variations in vocal pitch,
which may occur due to speaker differences or emo-
tional states. This augmentation captures variations
in speaker intonation while maintaining the original
speech content (Alex et al., 2023).
Noise Addition: Gaussian noise was added to the au-
dio with an amplitude factor of 0.005 to simulate en-
vironmental noise. This method increases robustness
by enabling the model to process noisy input data,
reflecting real-world recording conditions (Nugroho
et al., 2021).
Volume Adjustment: The amplitude of the audio
was scaled to 80% and 120% of its original level to
simulate variations in the recording conditions and
speaker distance from the microphone. These adjust-
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
668

ments ensure that the model can handle varying input
signal strengths (Zhou et al., 2017).
Audio Shifting: A random circular shift of up to
20% of audio length was applied to simulate mis-
alignments or varying speech onset times, enhancing
the model’s ability to handle such variances (Lounnas
et al., 2022).
Each augmented audio file was saved as a sepa-
rate sample, effectively increasing the dataset size and
introducing acoustic and temporal variability. This
augmentation strategy ensured that the model became
more resilient to natural variations and real-world
challenges. The augmented dataset was subsequently
used for model training, and the impact of each aug-
mentation type was evaluated for its contribution to
system performance.
3.4 Model Training and Evaluation
To assess the performance of our models, we com-
puted classification accuracy, precision, recall, and
F1-score. These metrics are standard in classifica-
tion tasks and provide insights into the model’s over-
all correctness (accuracy), the proportion of correctly
identified positive cases (precision), the ability to
identify all relevant positive cases (recall), and a bal-
anced measure of precision and recall (F1-score). De-
tailed definitions of these metrics are available in the
standard machine learning literature.
Figure 1: t-Sne visualization of speaker embeddings de-
rived from MFCC, formant, and pitch features across three
classes of speaking rates: L (Low), N (Normal), and S
(Speedy).
Figure 1 presents a t−SNE projection of the
speaker embeddings derived from multiple acoustic
features, including MFCCs, formant frequencies, and
pitch, across three distinct speaking rates. The visual
separation of the clusters demonstrates the discrim-
inative power of the combined feature set, particu-
larly in capturing the temporal variations in speech
signals. Notably, the dense overlap between some
points highlights the inherent challenges of speaker
verification under varying speaking rates. This result
reinforces the importance of integrating complemen-
tary features such as MFCCs and formants to enhance
the robustness of speaker verification systems, espe-
cially in real-world applications where speaking rate
variability is prevalent.
4 RESULTS AND DISCUSSION
We used 5-fold stratified cross-validation for model
training, running each fold for 50 epochs with a learn-
ing rate of 0.0001 and a batch size of 32. Early stop-
ping with a patience of 10 epochs was applied to pre-
vent overfitting and to retain optimal model weights.
The models were optimized using the Adam opti-
mizer and categorical cross-entropy loss for multi-
class classification tasks.
Table 1 presents the performance met-
rics—accuracy, precision, recall, and F1-score—of
different acoustic feature sets: MFCC, MFCC-delta-
delta, and formant frequencies (F0, F1–F4), evaluated
under augmented and non-augmented conditions.
Without Augmentation: The combined model
(MFCC, formant frequencies, and MFCC derivatives)
achieved the highest performance across all metrics,
with an accuracy of 70%, precision of 69%, recall of
67%, and F1-score of 68%. This result highlights the
advantage of combining diverse acoustic features to
capture speaker-specific information. In contrast, in-
dividual feature sets performed lower: MFCC-delta-
delta achieved 62% accuracy, MFCC alone reached
60%, and formant frequencies performed the weak-
est at 59%, underscoring their limited discriminatory
power when used in isolation.
With Augmentation: all models showed clear im-
provements, confirming its role in enhancing model
robustness. The combined model again outperformed
individual models, achieving 75% accuracy, with pre-
cision, recall, and F1-scores of 74%, 74%, and 73%,
respectively. Augmentation also notably improved in-
dividual feature sets. Formant frequencies exhibited a
significant increase in accuracy, rising to 68%, match-
ing the MFCC-delta-delta model, which improved to
67% accuracy. Even MFCC alone benefited, achiev-
ing 65% accuracy compared to 60% without augmen-
tation.
These results underscore the effectiveness of com-
bining spectral (MFCC), temporal (delta and delta-
delta), and source-related (formants) features to pro-
vide a comprehensive representation of speaker char-
acteristics. The combined model consistently out-
Speaker Verification Enhancement via Speaking Rate Dynamics in Persian Speechprints
669

Table 1: Model performance of MFCC, F0, and Formant Frequencies, and their combinations with and without augmentation.
Augmentation Model Accuracy Precision Recall F1-score
Without Aug.
MFCC 60% 61% 60% 59%
MFCC-delta-delta 62% 60% 63% 62%
F0, F1-F4 59% 56% 55% 55%
Combined Model 70% 69% 67% 68%
With Aug.
MFCC 65% 62% 62% 61%
MFCC-delta-delta 67% 64% 63% 64%
F0, F1-F4 68% 68% 67% 67%
Combined Model 75% 74% 74% 73%
performed individual feature sets in both conditions,
while data augmentation further enhanced model per-
formance, introducing variability that simulates real-
world speaking conditions. The superior results
achieved by the augmented combined model demon-
strate the robustness and generalizability of this multi-
feature approach, particularly in handling variations
in speaking styles and conditions, making it a strong
foundation for speaker verification systems.
5 DISCUSSION
The findings from this study highlight the impact of
combining multiple acoustic feature sets and apply-
ing data augmentation on the robustness and accu-
racy of speaker identification systems. Specifically,
the combined model (MFCC, MFCC-delta, MFCC-
delta-delta, and formant frequencies) outperformed
individual feature models in both augmented and
non-augmented conditions, achieving a 75% accuracy
with augmentation, compared to 70% without aug-
mentation. This confirms previous work suggesting
that hybrid approaches, which leverage both spectral
and temporal information, provide a more compre-
hensive representation of speaker-specific traits (De-
hak et al., 2010).
One reason for the success of MFCC-based fea-
tures in this context is their ability to capture speaker-
specific spectral envelopes, which have long been
considered the foundation for speaker recognition
tasks (Davis and Mermelstein, 1980). The delta and
delta-delta coefficients add temporal dynamics, al-
lowing the system to capture how the spectral proper-
ties change over time, which is particularly important
for handling variations in speaking rates and articula-
tion patterns. This finding aligns with previous find-
ings (Snyder et al., 2018; Choi et al., 2015), where in-
corporating temporal features significantly improved
speaker recognition performance under varied speak-
ing conditions. In this study, the inclusion of MFCC-
delta-delta improved accuracy from 60% to 62% in
non-augmented data, further demonstrating the im-
portance of modeling temporal fluctuations.
Formant frequencies, which represent the resonant
frequencies of the vocal tract, provided additional in-
formation that enhanced speaker identification when
combined with MFCCs, even though they underper-
formed as a standalone feature (59% accuracy with-
out augmentation). While formants capture impor-
tant physiological information about a speaker’s vocal
tract, they tend to be less reliable on their own, par-
ticularly when speech is affected by external factors
such as noise or varying speaking rates (Hansen and
Hasan, 2015). However, when paired with MFCCs,
formants contribute valuable vocal tract information,
improving the overall robustness of speaker recog-
nition systems. This echoes findings from Nath
and Kalita (Nath and Kalita, 2015), who demon-
strated that combining formants with other features
like MFCCs significantly enhanced speaker recog-
nition accuracy, with results nearing 100% in some
tasks. Similarly, Messaoud and Hamida (Messaoud
and Hamida, 2011) found that integrating formant fre-
quencies with MFCCs in a phone recognition system
reduced the phone error rate by 3%, further validating
the complementary nature of these features. These
studies highlight how formants and MFCCs work to-
gether by covering different aspects of the speech sig-
nal, making the combined approach highly effective
for speaker recognition.
Data augmentation played a pivotal role in en-
hancing system performance across all models. The
most significant improvements were observed in the
combined model, where accuracy increased from
70% to 75%, with similar gains in precision, recall,
and F1-score. This finding is consistent with prior re-
search, which demonstrated that augmentation meth-
ods such as time-stretching, pitch-shifting, and noise
addition increase the diversity of the training data,
enabling models to generalize better to unseen con-
ditions (Nugroho et al., 2021; Ko et al., 2015). In
our case, augmentation helped the model better han-
dle variations in speaking rate and background noise,
which are common in real-world applications. By ex-
posing the model to these variations during training,
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
670

we effectively reduced overfitting, thus improving its
ability to generalize to new data.
While MFCC-delta-delta and formant features
saw improvements with augmentation, MFCCs alone
also benefited, achieving a 65% accuracy compared
to 60% without augmentation. This confirms that
data augmentation is essential even when using ro-
bust features such as MFCCs, as it simulates real-
world variability, making the model more resilient
to changes in speech conditions (Koo et al., 2020).
The increased accuracy of formant frequencies (68%
with augmentation) suggests that, although formants
alone may struggle with speaker discrimination in
clean conditions, they become more useful when aug-
mented, as they help capture subtle articulatory varia-
tions that may emerge under different speaking envi-
ronments (Trottier et al., 2015).
5.1 Limitations and Future Work
A primary limitation of this study is the restricted
dataset size and its language specificity. The dataset,
consisting of only 18 male Persian speakers, limits the
generalizability of the findings. Speaker verification
systems often perform differently across languages
due to phonetic and prosodic variations, raising un-
certainty about the generalizability of these results
to other linguistic contexts. Additionally, the small
dataset may have constrained the model’s ability to
capture broader speaker variability. A more extensive
and diverse dataset would be necessary to assess the
system’s robustness on a larger scale.
This study underscores the value of a hybrid
acoustic feature approach for speaker identification,
particularly when combined with data augmentation.
The integration of MFCCs, delta features, and for-
mant frequencies provided a multi-dimensional rep-
resentation of vocal traits, enhancing performance un-
der varied conditions. Future research could investi-
gate additional feature combinations, such as prosodic
features (e.g., intonation, rhythm) and voice qual-
ity measures (e.g., jitter, shimmer), to improve ro-
bustness, particularly for emotional speech or non-
standard speaking styles.
While this study focused on traditional feature
extraction methods, the use of deep learning-based
embeddings—such as x-vectors(Snyder et al., 2018)
or Transformer-based models(Vaswani, 2017)—holds
significant potential. These models can learn speaker-
specific characteristics directly from raw audio, re-
ducing the reliance on manual feature engineering.
Combining such approaches with advanced augmen-
tation techniques could further enhance performance,
enabling speaker identification systems to handle
challenging real-world conditions, such as noisy envi-
ronments, emotional variability, and diverse speaking
styles.
6 CONCLUSION
In this study, we investigated the resilience of vari-
ous acoustic features in speaker identification across
different speaking rates. The findings reveal a hierar-
chy of effectiveness among the examined parameters.
The results indicate that vowel formant frequencies
demonstrate a degree of resilience against changes in
speaking rate, achieving an accuracy of 80 in speaker
identification tasks. This suggests that while for-
mant frequencies capture relevant speaker-specific in-
formation, their ability to distinguish between speak-
ers may be somewhat compromised when faced with
variations in speaking rate. Given the strong corre-
lation between formant frequencies and the invariant
physiological dimensions of the vocal tract, it is un-
surprising that these frequencies remain relatively sta-
ble across varying speech rates. Despite the relatively
stable vocal tract, dynamic articulatory adjustments
required for different speaking rates could potentially
introduce variability into formant measurements.
REFERENCES
Agwuele, A., Sussman, H. M., and Lindblom, B. (2009).
The effect of speaking rate onconsonant vowel coar-
ticulation. Phonetica, 65(4):194–209.
Alex, A., Wang, L., Gastaldo, P., and Cavallaro, A. (2023).
Data augmentation for speech separation. Speech
Commun., 152:102949.
Anupama, V., Amrutha, C., Varshini, G. A., Nandan, G.
S. G., and Vivek, G. S. S. (2022). A mfcc-cnn based
voice authentication security. Int. J. Eng. Technol.
Manag. Sci., 4(6).
Asadi, H., Nourbakhsh, M., Sasani, F., and Dellwo, V.
(2018). Examining long-term formant frequency as a
forensic cue for speaker identification: An experiment
on persian. In Proc. 1st Int. Conf. Lab. Phon. Phonol.,
pages 21–28.
Bahari, M. H. and Van Hamme, H. (2011). Speaker age
estimation and gender detection based on supervised
non-negative matrix factorization. In BIOMS, pages
1–6. IEEE.
Boersma, P. and Weenink, D. (2021). Praat: Doing phonet-
ics by computer [computer program](version 6.2. 22).
Retrieved fromwww. praat. org.
Chakravarty, N. and Dua, M. (2023). Data augmentation
and hybrid feature amalgamation to detect audio deep
fake attacks. Physica Scripta, 98.
Choi, Y. H., Ban, S. M., Kim, K.-W., and Kim, H. S.
(2015). Evaluation of frequency warping based fea-
Speaker Verification Enhancement via Speaking Rate Dynamics in Persian Speechprints
671

tures and spectro-temporal features for speaker recog-
nition. Phonetics and Speech Sciences, 7(1):3–10.
Corretge, R. (2022). Praat vocal toolkit. Available at: http:
//www.praatvocaltoolkit.com.
Davis, S. and Mermelstein, P. (1980). Comparison of para-
metric representations for monosyllabic word recog-
nition in continuously spoken sentences. IEEE Trans.
Acoust. Speech Signal Process., 28(4):357–366.
Dehak, N., Kenny, P. J., Dehak, R., Dumouchel, P., and
Ouellet, P. (2010). Front-end factor analysis for
speaker verification. IEEE Trans. Audio Speech Lang.
Process., 19(4):788–798.
Gay, T., Ushijima, T., Hiroset, H., and Cooper, F. S. (1974).
Effect of speaking rate on labial consonant-vowel ar-
ticulation. Journal of Phonetics, 2(1):47–63.
Gold, E., French, P., and Harrison, P. (2013). Examining
long-term formant distributions as a discriminant in
forensic speaker comparisons under a likelihood ratio
framework. In Proc. Meet. Acoust., volume 19. AIP
Publishing.
Hansen, J. H. and Hasan, T. (2015). Speaker recognition by
machines and humans: A tutorial review. IEEE Signal
Process. Mag., 32(6):74–99.
Hinton, G., Deng, L., Yu, D., Dahl, G. E., Mohamed, A.-
r., Jaitly, N., Senior, A., Vanhoucke, V., Nguyen, P.,
Sainath, T. N., et al. (2012). Deep neural networks for
acoustic modeling in speech recognition: The shared
views of four research groups. IEEE Signal Process.
Mag., 29(6):82–97.
Imaizumi, S. and Kiritani, S. (1989). Effect of speaking
rate on formant trajectories and inter-speaker varia-
tions. Ann. Bull. RILP, 23:27–37.
Jahangir, R., Teh, Y. W., Memon, N. A., Mujtaba, G., Za-
reei, M., Ishtiaq, U., Akhtar, M. Z., and Ali, I. (2020).
Text-independent speaker identification through fea-
ture fusion and deep neural network. IEEE Access,
8:32187–32202.
Ko, T., Peddinti, V., Povey, D., and Khudanpur, S. (2015).
Audio augmentation for speech recognition. In Inter-
speech, volume 2015, page 3586.
Koo, H., Jeong, S., Yoon, S., and Kim, W. (2020). Develop-
ment of speech emotion recognition algorithm using
mfcc and prosody. ICEIC, pages 1–4.
Liu, X., Sahidullah, M., and Kinnunen, T. (2021). Learn-
able mfccs for speaker verification. In ISCAS, pages
1–5. IEEE.
Lounnas, K., Lichouri, M., and Abbas, M. (2022). Analysis
of the effect of audio data augmentation techniques
on phone digit recognition for algerian arabic dialect.
ICAASE, pages 1–5.
McDougall, K. (2006). Dynamic features of speech and the
characterization of speakers: Toward a new approach
using formant frequencies. Int. J. Speech Lang. Law,
13(1):89–126.
McFee, B., Raffel, C., Liang, D., Ellis, D. P., McVicar, M.,
Battenberg, E., and Nieto, O. (2015). librosa: Audio
and music signal analysis in python. In SciPy, pages
18–24.
Mefferd, A. S. and Green, J. R. (2010). Articulatory-to-
acoustic relations in response to speaking rate and
loudness manipulations. J. Speech Lang. Hear. Res.,
53:1206–1219.
Messaoud, Z. B. and Hamida, A. (2011). Combining for-
mant frequency based on variable order lpc coding
with acoustic features for timit phone recognition. Int.
J. Speech Technol., 14:393.
Nath, D. and Kalita, S. (2015). Composite feature selection
method based on spoken word and speaker recogni-
tion. Int. J. Comput. Appl., 121:18–23.
Nolan, F. (1987). The phonetic bases of speaker recogni-
tion: Cambridge studies in speech science and com-
munication, cambridge university press, cambridge,
1983, 221 pp. isbn 0-521-24486-2.
Nugroho, K., Noersasongko, E., Purwanto, Muljono, and
Setiadi, D. (2021). Enhanced indonesian ethnic
speaker recognition using data augmentation deep
neural network. J. King Saud Univ. Comput. Inf. Sci.,
34:4375–4384.
Reynolds, D. A., Quatieri, T. F., and Dunn, R. B. (2000).
Speaker verification using adapted gaussian mixture
models. Digital signal processing, 10(1-3):19–41.
Rose, P. (2002). Forensic Speaker Identification. Interna-
tional Forensic Science and Investigation. Taylor &
Francis.
Shahrebabaki, A. S., Imran, A. S., Olfati, N., and Svendsen,
T. (2018). Acoustic feature comparison for different
speaking rates. In Proc. Human-Computer Interaction
(HCI), pages 176–189. Springer.
Shaiman, S., Adams, S. G., and Kimelman, M. D. (1997).
Velocity profiles of lip protrusion across changes in
speaking rate. J. Speech Lang. Hear. Res., 40(1):144–
158.
Snyder, D., Garcia-Romero, D., Sell, G., Povey, D., and
Khudanpur, S. (2018). X-vectors: Robust dnn embed-
dings for speaker recognition. ICASSP 2018, pages
5329–5333.
Tilsen, S. (2014). Selection and coordination of articulatory
gestures in temporally constrained production. Jour-
nal of Phonetics, 44:26–46.
Trottier, L., Chaib-draa, B., and Giguere, P. (2015). Tempo-
ral feature selection for noisy speech recognition. In
Proc. Can. Conf. Artif. Intell., pages 155–166.
Tuller, B., Harris, K. S., and Kelso, J. S. (1982). Stress and
rate: Differential transformations of articulation. J.
Acoust. Soc. Am., 71(6):1534–1543.
Vaswani, A. (2017). Attention is all you need. Advances in
Neural Information Processing Systems.
Weismer, G. and Berry, J. (2003). Effects of speaking rate
on second formant trajectories of selected vocalic nu-
clei. J. Acoust. Soc. Am., 113(6):3362–3378.
Xie, W., Nagrani, A., Chung, J. S., and Zisserman,
A. (2019). Utterance-level aggregation for speaker
recognition in the wild. In ICASSP 2019, pages 5791–
5795. IEEE.
Zeng, X., Yin, S., and Wang, D. (2015). Learning
speech rate in speech recognition. arXiv preprint
arXiv:1506.00799.
Zhou, Y., Xiong, C., and Socher, R. (2017). Improved reg-
ularization techniques for end-to-end speech recogni-
tion. ArXiv, abs/1712.07108.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
672