Ardelia, Danika Najwa (2026) Classification of Javanese Dialects using A Hybrid CNN-TCN Deep Learning Architecture. Undergraduate thesis, UPN Veteran Jawa Timur.
|
Text (Cover)
22081010103.-cover.pdf Download (1MB) | Preview |
|
|
Text (Bab 1)
22081010103.-bab1.pdf Download (260kB) | Preview |
|
![[img]](https://repository.upnjatim.ac.id/style/images/fileicons/text.png) |
Text (Bab 2)
22081010103.-bab2.pdf Restricted to Repository staff only until 30 June 2029. Download (799kB) |
|
|
Text (Bab 3)
22081010103.-bab3.pdf Restricted to Repository staff only until 30 June 2029. Download (1MB) |
|
|
Text (Bab 4)
22081010103.-bab4.pdf Restricted to Repository staff only until 30 June 2029. Download (2MB) |
|
|
Text (Bab 5)
22081010103.-bab5.pdf Download (188kB) | Preview |
|
|
Text (Daftar Pustaka)
22081010103.-daftarpustaka.pdf Download (232kB) | Preview |
|
|
Text (Lampiran)
22081010103.-lampiran.pdf Restricted to Repository staff only until 30 June 2029. Download (2MB) |
Abstract
The decline in the use of regional languages in Indonesia, including Javanese, threatens the preservation of local cultural identity. Classifying Javanese dialects is challenging due to subtle variations in pitch and tempo. This study proposes a hybrid Convolutional Neural Network (CNN) and Temporal Convolutional Network (TCN) architecture to identify five Javanese dialects: Arekan, Pantura (Banten), Banyumas, Cirebon, and Solo-Yogya. The CNN front-end extracts spatial features from Log-Mel Spectrograms, while the TCN back-end models long-term temporal dependencies of the speech signals. Data was collected from The Language Archive (TLA) and YouTube, then preprocessed through 16 kHz resampling, Voice Activity Detection (VAD), 10-second segmentation with 30% overlap, and feature extraction. Noise Injection and SpecAugment techniques were applied to enrich the data and prevent overfitting. The system was evaluated using Stratified Group 5-Fold Cross Validation, employing a Soft Voting probability aggregation mechanism to combine segmentlevel predictions into a final audio-level decision. The results demonstrate that the CNN-TCN integration achieved an accuracy of 90.68% and an F1-Score of 0.9032. This hybrid architecture significantly outperformed the single CNN model (83.53%) and TCN model (86.37%). Furthermore, it effectively resolved cross-dialect classification confusion, notably improving the recognition of the Arekan dialect. Finally, the optimal classification model was successfully implemented into an interactive web-based application interface to facilitate dialect identification for end users.
| Item Type: | Thesis (Undergraduate) | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
||||||||||||
| Subjects: | Q Science > QA Mathematics > QA76.87 Neural computers | ||||||||||||
| Divisions: | Faculty of Computer Science > Departemen of Informatics | ||||||||||||
| Depositing User: | Danika Najwa Ardelia | ||||||||||||
| Date Deposited: | 30 Jun 2026 06:34 | ||||||||||||
| Last Modified: | 30 Jun 2026 06:34 | ||||||||||||
| URI: | https://repository.upnjatim.ac.id/id/eprint/54311 |
Actions (login required)
 |
View Item |