TY - GEN
T1 - Multiple Deep Neural Network Models for Facial Motion to Vowel Recognition for Bahasa Indonesia
T2 - 6th International Conference on Computer Engineering, Network and Intelligent Multimedia, CENIM 2025
AU - Rahman, Muhammad Daffa Abiyyu
AU - Wicaksono, Alif Aditya
AU - Yuniarno, Eko Mulyanto
AU - Nugroho, Supeno Mardi Susiki
AU - Suryo, Arief Trisno Eko
AU - Cahyono, Gunawan Rudi
N1 - Publisher Copyright:
©2025 IEEE.
PY - 2025
Y1 - 2025
N2 - —Detection of facial motion to determine the vowel being spoken by a subject in Bahasa Indonesia needs to be explored in depth. A correlation was found between facial movements and the acoustic sound produced indicates that it is possible to determine the letters spoken by an individual based on their facial motion. Utilizing landmark, convolutional neural network and time-distribution models, it is possible to read a sequence of landmark from images to determine vowel being spoken. However, selecting the right model can be a challenge, due to the complexity of speaking as an action and the vast array of options available. As such, five different models are compared in this research, consisting of CNN2D, CNN-LSTM, CNN-BiLSTM, CNN-GRU, and CNN-BiGRU models to find the best performing model. After a deep analysis, CNN-BiLSTM model performed best at 72.13% accuracy and CNN-BiGRU and CNN2D as alternatives.
AB - —Detection of facial motion to determine the vowel being spoken by a subject in Bahasa Indonesia needs to be explored in depth. A correlation was found between facial movements and the acoustic sound produced indicates that it is possible to determine the letters spoken by an individual based on their facial motion. Utilizing landmark, convolutional neural network and time-distribution models, it is possible to read a sequence of landmark from images to determine vowel being spoken. However, selecting the right model can be a challenge, due to the complexity of speaking as an action and the vast array of options available. As such, five different models are compared in this research, consisting of CNN2D, CNN-LSTM, CNN-BiLSTM, CNN-GRU, and CNN-BiGRU models to find the best performing model. After a deep analysis, CNN-BiLSTM model performed best at 72.13% accuracy and CNN-BiGRU and CNN2D as alternatives.
KW - Bahasa Indonesia
KW - Deep Learning
KW - Facial Motion
KW - Vowel Recognition
UR - https://www.scopus.com/pages/publications/105033158520
U2 - 10.1109/CENIM67940.2025.11326198
DO - 10.1109/CENIM67940.2025.11326198
M3 - Conference contribution
AN - SCOPUS:105033158520
T3 - Proceeding of the International Conference on Computer Engineering, Network and Intelligent Multimedia 2025, CENIM 2025
SP - 171
EP - 178
BT - Proceeding of the International Conference on Computer Engineering, Network and Intelligent Multimedia 2025, CENIM 2025
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 25 November 2025 through 26 November 2025
ER -