Jovan Galić

Prof. Dr. Jovan Galić

orcid

Research Focus & Metrics

Academic Career Length

9Years (2018-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
12 September 2024 in Applied Sciences
Verified
Journal Article
Open access
PDF file
JGJovan Galić
Jovan Galić
BMBranko Marković
Branko Marković
ĐGĐorđe Grozdić
Đorđe Grozdić
Branislav PopovićBranislav Popović
Branislav Popović
SSSlavko Sajic
Slavko Sajic
Modern Automatic Speech Recognition (ASR) systems are primarily designed to recognize normal speech. Due to a considerable acoustic mismatch between normal speech and whisper, ASR systems suffer from a significant loss of performance in whisper recognition. Creating large databases of whispered speech is expensive and time-consuming, so research studies explore the synthetic generation using pre-existing normal or whispered speech databases. The impact of standard audio data augmentation techniques on the accuracy of isolated-word recognizers based on Hidden Markov Models (HMM) and Convolutional Neural Networks (CNN) is examined in this research study. Furthermore, the study explores the potential of inverse filtering as an augmentation strategy for producing pseudo-whisper speech. The Whi-Spe speech database, containing recordings in normal and whisper phonation, is utilized for data augmentation, while the internally recorded speech database, developed specifically for this study, is employed for testing purposes. Experimental results demonstrate statistically significant improvement in performance when employing data augmentation strategies and inverse filtering.

25 November 2025
Verified
Conference Paper
Closed access
BMBranko Marković
Branko Marković
VLVeljko Lončarević
Veljko Lončarević
JGJovan Galić
Jovan Galić
This paper presents the implementation of 1DCNN for whispered speech recognition. At the front end, MFCCs were used to construct feature vectors, while the back end employed a 1D-CNN for training, validation, and testing. Speech patterns were taken from the Whi-Spe database, considering two speakers (one male and one female). Four types of feature vectors were examined through different scenarios: Normal/Normal, Whisper/Whisper, Normal/Whisper, and Whisper/Normal. The results confirm the effectiveness of 1DCNN for speech recognition.


1 June 2018 in SPIIRAS Proceedings
Verified
Journal Article
Open access
PDF file
JGJovan Galić
Jovan Galić
SJSlobodan Toma Jovičić
Slobodan Toma Jovičić
Vlado DelićVlado Delić
Vlado Delić
BMBranko Rade Marković
Branko Rade Marković
DPDragana Staniša Šumarac Pavlović
Dragana Staniša Šumarac Pavlović
ĐGĐorđe Tomislav Grozdić
Đorđe Tomislav Grozdić
Отсутствие достаточного количества данных шепотной речи для обучения является серьезной проблемой для современных систем автоматического распознавания речи (АРР). Из-за большого акустического различия между обычной и шепотной речью АРР системы значительно снижают производительность при обработке шепота. В статье приведен анализ подходов к распознаванию нейтральной и шепотной речи на основе традиционных скрытых марковских моделей (СММ) для дикторозависимых (SD) и дикторонезависимых (SI) случаев. Особое внимание уделяется распознаванию шепотной речи с использованием нейтральной речи на этапе обучения (сценарий N/W). Система АРР разработана для распознавания изолированных слов из базы данных (Whi-Spe), включающей пары слов реально произнесенной речи нейтрально и шепотом. В сценарии N/W увеличение надежности достигается с применением предлагаемого частотного преобразования, изначально разработанного для сжатия и декомпрессии речевого сигнала в цифровых телекоммуникационных системах. Вместе с тем сохраняются хорошие показатели в распознавании нейтральной речи. По сравнению с базовой моделью распознавания с применением Мел-частотных кепстральных коэффициентов (MFCC) точность распознавания слов с использованием кепстральных коэффициентов, полученных с помощью предложенного частотного деформирования (обозначаемого как μFCC), улучшена на 7,36% (SD) и 3,44% (SI) в абсолютных значениях. Кроме того, F-мера (гармоническое среднее значение точности и полноты) для векторов признаков μFCC увеличивается на 6,90% (SD) и 3,59 %(SI). Статистические тесты подтверждают значимость достигнутого улучшения точности распознавания.

23 September 2026 in Circuits, Systems, and Signal Processing
Verified
Journal Article
Open access
PDF file
BMBranko Marković
Branko Marković
JGJovan Galić
Jovan Galić
AKAsutosh Kar
Asutosh Kar
Whispered speech poses persistent challenges for automatic speech recognition (ASR) owing to the absence of vocal fold vibration and the dominance of turbulent airflow excitation, which substantially alters spectral and temporal energy distributions. Conventional front-end features, such as Perceptual Linear Predictive Cepstral Coefficients (PLPCC), exhibit limited robustness in these conditions, particularly under cross-mode (mismatched) training–testing scenarios. To address this, we propose Teager Energy–based Perceptual Linear Predictive Cepstral Coefficients (TEPLPCC), a novel feature extraction framework that integrates the nonlinear Teager Energy Operator (TEO) within the perceptually motivated PLP analysis. The proposed method exploits instantaneous energy variations to characterize the nonlinear aerodynamic behavior inherent in whispered speech. Extensive experiments were conducted on the Whi-Spe bimodal (normal/whisper) speech database using three recognition back ends—Dynamic Time Warping (DTW), Hidden Markov Models (HMM), and Convolutional Neural Networks (CNN)—under both matched (Normal/Normal, Whisper/Whisper) and mismatched (Normal/Whisper, Whisper/Normal) conditions. Feature normalization via Cepstral Mean Subtraction (CMS) was employed to mitigate channel and mode variability. Results demonstrate that TEPLPCC consistently surpasses baseline PLPCC and existing TEO-derived features (TELFCC and TEMFCC), achieving up to 13% relative improvement in mismatched conditions and approximately 10% higher recognition accuracy on average. These outcomes establish TEPLPCC as a perceptually consistent and energy-aware representation, significantly advancing the robustness of ASR systems for whispered speech.

1 January 2025 in Serbian Journal of Electrical Engineering
Verified
Journal Article
Open access
PDF file
BMBranko Marković
Branko Marković
VLVeljko Loncarevic
Veljko Loncarevic
JGJovan Galić
Jovan Galić
Discrete Wavelet Transforms (DWTs) provide time?frequency representations that are well suited for nonstationary signals such as speech. This study presents a comparison of four wavelet families (Daubechies, Symlets, Coiflets, and Biorthogonal) for bimodal automatic speech recognition across two speech modes (normal and whispered). Experiments use the Whi-Spe database comprising ten speakers (five female and five male). A Dynamic Time Warping (DTW) back-end performs sequence alignment and recognition. Results are reported via summary tables, histograms, and confusion matrices and reveal systematic differences among the wavelet families, identifying the most effective transform for bimodal recognition. These findings provide practical guidance for selecting wavelet-based front ends in whisper-robust automatic speech recognition (ASR) systems.