prikaz prve stranice dokumenta Prepoznavanje govornika korištenjem dubokih neuronskih mreža
  Download
PDF 3.16 MB
master's thesis
Prepoznavanje govornika korištenjem dubokih neuronskih mreža
2024. urn:nbn:hr:200:455092
Markić, Luka
Josip Juraj Strossmayer University of Osijek
Faculty of Electrical Engineering, Computer Science and Information Technology Osijek
Department of Computer Engineering and Automation
Chair of Automation and Robotics

Cite this document

Markić, L. (2024). Prepoznavanje govornika korištenjem dubokih neuronskih mreža (Master's thesis). Osijek: Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek. Retrieved from https://urn.nsk.hr/urn:nbn:hr:200:455092

Markić, Luka. "Prepoznavanje govornika korištenjem dubokih neuronskih mreža." Master's thesis, Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, 2024. https://urn.nsk.hr/urn:nbn:hr:200:455092

Markić, Luka. "Prepoznavanje govornika korištenjem dubokih neuronskih mreža." Master's thesis, Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, 2024. https://urn.nsk.hr/urn:nbn:hr:200:455092

Markić, L. (2024). 'Prepoznavanje govornika korištenjem dubokih neuronskih mreža', Master's thesis, Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, accessed 22 January 2025, https://urn.nsk.hr/urn:nbn:hr:200:455092

Markić L. Prepoznavanje govornika korištenjem dubokih neuronskih mreža [Master's thesis]. Osijek: Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek; 2024 [cited 2025 January 22] Available at: https://urn.nsk.hr/urn:nbn:hr:200:455092

L. Markić, "Prepoznavanje govornika korištenjem dubokih neuronskih mreža", Master's thesis, Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, Osijek, 2024. Available at: https://urn.nsk.hr/urn:nbn:hr:200:455092

Please login to the repository to save this object to your list.
Metadata
TitlePrepoznavanje govornika korištenjem dubokih neuronskih mreža
Title (english)Speaker recognition using deep neural networks
AuthorLuka Markić
MentorRatko Grbić (mentor)
Committee memberRatko Grbić (predsjednik povjerenstva)
Committee memberRobert Cupec (član povjerenstva)
Committee memberPetra Pejić (član povjerenstva)
GranterJosip Juraj Strossmayer University of Osijek
Faculty of Electrical Engineering, Computer Science and Information Technology Osijek
(Department of Computer Engineering and Automation)
(Chair of Automation and Robotics)
Osijek
Defense date and country2024-10-01, Croatia
Scientific / art field,
discipline and subdiscipline		TECHNICAL SCIENCES
Computing
Artificial Intelligence
Abstract
U okviru ovog rada predstavljen je način rješavanja zadatka verifikacije govornika primjenom dubokih neuronskih mreža. Prvotno su definirani zadaci identifikacije i verifikacije govornika te su predstavljena moderna rješenja za rješavanje tih problema, pri čemu se ističu rješenja ostvarena korištenjem konvolucijske i sijamske neuronske mreže. Ovim radom zadatak verifikacije govornika se rješava treniranjem sijamske neuronske mreže na temelju logaritamskih Mel spektrograma stvorenih iz zvučnih zapisa GRID skupa podataka. Primjenom augmentacija Gaussovog šuma i amplitudne augmentacije spektrograma na trening skupu podataka treniraju se nove mreže kojima se procjenjuje utjecaj augmentacija na rad verifikacije govornika. S ciljem povećanja raznolikosti skupa podataka, stvoren je vlastiti skup podataka s 12 različitih govornika. Korištenje tog skupa i procesom prijenosnog učenja na temelju već istreniranih mreža na GRID skupu podataka dobivene su nove mreže. Nadalje, izrađena je aplikacija koja omogućuje primjenu istreniranih mreža za verifikaciju govornika u stvarnom okruženju. U konačnici, sve izgrađene mrže su detaljno evaluirane na testnom skupu podataka.
Abstract (english)
This master’s thesis presents a method for solving the speaker verification task by using deep neural networks. Firstly, the tasks of speaker identification and verification are defined, followed by presenting modern solutions for these problems, with a focus on solutions achieved using Convolutional Neural Networks and Siamese Neural Networks. In this paper, the speaker verification task is solved by training a Siamese Neural Network based on log-Mel spectrograms created on audio recordings of the GRID dataset. By applying data augmentations on training dataset, such as Gaussian noise and amplitude augmentation of spectrograms, new networks are trained to evaluate the impact of these augmentations on speaker verification performance. To increase the diversity of the GRID dataset, a custom dataset with 12 different speakers was created. Using this dataset and a transfer learning process based on networks already trained on the GRID dataset, new networks were trained. Furthermore, a desktop application was developed that enables the use of trained networks for speaker verification in real-world environment. In the end, all constructed networks were thoroughly evaluated on the test dataset.
Keywords
Keywords (english)
Languagecroatian
URN:NBNurn:nbn:hr:200:455092
Study programmeTitle: Graduate University Study Programme in Computer Engineering
Study programme type: university
Study level: graduate
Academic / professional title: sveučilišni/a magistar/magistra inženjer/inženjerka računarstva (sveučilišni/a magistar/magistra inženjer/inženjerka računarstva)
Type of resourceText
File originBorn digital
Access conditionsOpen access
Terms of useLicence In copyright icon
Created on2024-10-04 11:04:48