Gunarso Gunarso
Gunarso Gunarso has a total of 34 co-authors. Here are the top co-authors by number of shared publications:
- Mohammad Teduh Uliniansyah
National Research and Innovation Agency (BRIN)
6 shared publications - Agung Santosa
National Research and Innovation Agency (BRIN)
5 shared publications - Asril Jarin
National Research and Innovation Agency (BRIN)
5 shared publications - Elvira Nurfadhilah
Brin, Indonesia
4 shared publications - Andi Djalal Latief
National Research and Innovation Agency (BRIN), Indonesia
3 shared publications - Siska Pebiana
National Research and Innovation Agency, Indonesia
3 shared publications
Latest Publications
- Open Access
- 1 October 2026
- Published by Institute of Advanced Engineering and Science in IAES International Journal of Artificial Intelligence (IJ-AI)
- Vol. 15 (5) , 4893-4906
Empirical evaluation of generative fusion decoding for Indonesian general and medical speech recognition
This study empirically evaluates generative fusion decoding (GFD) with open-weight large language models for Indonesian automatic speech recognition (ASR) under two controlled offline conditions: general read speech (FLEURS-ID) and medical dictation (Harkit). The training-free byte-level fusion integrates three frozen language-model backends—SEA-LION, Sahabat-AI, and Mistral—with frozen Whisper-large-v3 during decoding. A development partition was used to select the static fusion weight and prompting condition, and a held-out test partition was used for final evaluation. On FLEURS-ID, the best GFD configuration reduced corpus-level word error rate (WER) from 5.91% to 4.17%, corresponding to a 29.36% relative reduction. On Harkit, it reduced WER from 15.89% to 12.30%, a 22.59% relative reduction. Paired utterance-level tests yielded p < 0.01 for both datasets. Several non-prompted configurations already improved over the unfused Whisper-large-v3 baseline, suggesting that the external language-model prior contributed more consistently than explicit prompting. Prompting produced limited changes on FLEURS-ID and larger but backend-dependent changes on Harkit. The results support GFD as a training-free decoding strategy under the evaluated conditions, while broader validation is required for noisy, spontaneous, streaming, and clinical deployment settings.
- 1 October 2026
- Published by Institute of Advanced Engineering and Science in IAES International Journal of Artificial Intelligence (IJ-AI)
- Vol. 15 (5) , 4893-4906
Open AccessEmpirical evaluation of generative fusion decoding for Indonesian general and medical speech recognition
This study empirically evaluates generative fusion decoding (GFD) with open-weight large language models for Indonesian automatic speech recognition (ASR) under two controlled offline conditions: general read speech (FLEURS-ID) and medical dictation (Harkit). The training-free byte-level fusion integrates three frozen language-model backends—SEA-LION, Sahabat-AI, and Mistral—with frozen Whisper-large-v3 during decoding. A development partition was used to select the static fusion weight and prompting condition, and a held-out test partition was used for final evaluation. On FLEURS-ID, the best GFD configuration reduced corpus-level word error rate (WER) from 5.91% to 4.17%, corresponding to a 29.36% relative reduction. On Harkit, it reduced WER from 15.89% to 12.30%, a 22.59% relative reduction. Paired utterance-level tests yielded p < 0.01 for both datasets. Several non-prompted configurations already improved over the unfused Whisper-large-v3 baseline, suggesting that the external language-model prior contributed more consistently than explicit prompting. Prompting produced limited changes on FLEURS-ID and larger but backend-dependent changes on Harkit. The results support GFD as a training-free decoding strategy under the evaluated conditions, while broader validation is required for noisy, spontaneous, streaming, and clinical deployment settings.
- 30 June 2026
- Published by IGI Global Scientific Publishing
- p. 53-88
Safety and Reliability Evaluation of Medical Chatbots Powered by Large Language Models
LLMs have new probabilistic safety problems that cannot be detected with the current benchmark measures of accuracy such as hallucination, automation bias, demographic unfairness, and privacy. In this chapter, we propose SafeMed, which is a multi-dimensional assessment framework consisting of ten dimensions for clinical safety evaluation in LLMs. SafeMed considers the following aspects: clinical accuracy, hallucination control, robustness, calibration, bias and fairness, ethics, transparency, privacy, safety of human-automation interaction, and operational feasibility. SafeMed is derived based on global standards such as WHO Guidelines for AI Ethics, EU AI Act, and FDA Good Machine Learning Practices (GMLP), and it uses Indonesian regulations as a case study involving ChatGPT-based clinical assistants, Med-PaLM, and RAG. Adversarial testing and human-in-the-loop oversight are proposed as minimum deployment standards. The chapter concludes with a practical evaluation pipeline and governance recommendations for clinicians, administrators, and policymakers.
- 30 June 2026
- Published by IGI Global Scientific Publishing
- p. 53-88
Safety and Reliability Evaluation of Medical Chatbots Powered by Large Language Models
LLMs have new probabilistic safety problems that cannot be detected with the current benchmark measures of accuracy such as hallucination, automation bias, demographic unfairness, and privacy. In this chapter, we propose SafeMed, which is a multi-dimensional assessment framework consisting of ten dimensions for clinical safety evaluation in LLMs. SafeMed considers the following aspects: clinical accuracy, hallucination control, robustness, calibration, bias and fairness, ethics, transparency, privacy, safety of human-automation interaction, and operational feasibility. SafeMed is derived based on global standards such as WHO Guidelines for AI Ethics, EU AI Act, and FDA Good Machine Learning Practices (GMLP), and it uses Indonesian regulations as a case study involving ChatGPT-based clinical assistants, Med-PaLM, and RAG. Adversarial testing and human-in-the-loop oversight are proposed as minimum deployment standards. The chapter concludes with a practical evaluation pipeline and governance recommendations for clinicians, administrators, and policymakers.
- Open Access
- 13 September 2025
- Published by Universitas Airlangga in Journal of Aquaculture and Fish Health
- Vol. 14 (3) , 321-335
Sustainability Status of Vannamei Shrimp (Litopenaeus vannamei) Aquaculture in Millennial Pond Model
Vannamei shrimp (Litopenaeus vannamei) aquaculture activities in Jepara Regency are conducted at Brackish Water Aquaculture Centers (Balai Besar Perikanan Budidaya Air Payau/BBPBAP) Jepara with a millennial model. This aquaculture model features a critical point in the middle of the pond, which aims to reduce the risk of failure due to pathogenic virus attacks. However, in running a vannamei shrimp aquaculture, an appropriate strategy is required for business development and sustainability. In this study, a multidimensional scaling (MDS) method was used to analyze the ecological, economic, social, technological, and legal-institutional dimensions. The study was conducted during October-December 2024, in which primary data collection was carried out through interviews with respondents and secondary data from related data references. The results of the study showed that the sustainability values were 88.39 (ecological dimension), 81.05 (economic dimension), 75.25 (social dimension), 99.996 (technological dimension), and 99.996 (legal-institutional dimension). The average value of the sustainability index of vannamei shrimp farming in the millennial pond of BBPBAP Jepara was very sustainable, with a score of 88.94. Based on the leverage analysis, sensitive attributes obtained in this study were ecological domain (1) disease attacks (2) land utilization level (3) land suitability, economic domain (1) price trends, (2) market reach, (3) cultivation profits, social domain (1) education level, (2) community involvement/labor absorption, (3) availability of transportation and communication facilities, technology domain (1) pond construction (2) application of biosecurity and legal domain & (1) institutional training activities and socialization about cultivation businesses (2) availability of cultivator groups.
- 13 September 2025
- Published by Universitas Airlangga in Journal of Aquaculture and Fish Health
- Vol. 14 (3) , 321-335
Open AccessSustainability Status of Vannamei Shrimp (Litopenaeus vannamei) Aquaculture in Millennial Pond Model
Vannamei shrimp (Litopenaeus vannamei) aquaculture activities in Jepara Regency are conducted at Brackish Water Aquaculture Centers (Balai Besar Perikanan Budidaya Air Payau/BBPBAP) Jepara with a millennial model. This aquaculture model features a critical point in the middle of the pond, which aims to reduce the risk of failure due to pathogenic virus attacks. However, in running a vannamei shrimp aquaculture, an appropriate strategy is required for business development and sustainability. In this study, a multidimensional scaling (MDS) method was used to analyze the ecological, economic, social, technological, and legal-institutional dimensions. The study was conducted during October-December 2024, in which primary data collection was carried out through interviews with respondents and secondary data from related data references. The results of the study showed that the sustainability values were 88.39 (ecological dimension), 81.05 (economic dimension), 75.25 (social dimension), 99.996 (technological dimension), and 99.996 (legal-institutional dimension). The average value of the sustainability index of vannamei shrimp farming in the millennial pond of BBPBAP Jepara was very sustainable, with a score of 88.94. Based on the leverage analysis, sensitive attributes obtained in this study were ecological domain (1) disease attacks (2) land utilization level (3) land suitability, economic domain (1) price trends, (2) market reach, (3) cultivation profits, social domain (1) education level, (2) community involvement/labor absorption, (3) availability of transportation and communication facilities, technology domain (1) pond construction (2) application of biosecurity and legal domain & (1) institutional training activities and socialization about cultivation businesses (2) availability of cultivator groups.
- Open Access
- 1 June 2025
- Published by Institute of Advanced Engineering and Science in IAES International Journal of Artificial Intelligence (IJ-AI)
- Vol. 14 (3) , 2389-2401
Modeling sentiment analysis of Indonesian biodiversity policy Tweets using IndoBERTweet
This study develops and evaluates a sentiment analysis model using IndoBERTweet to analyze Twitter data on Indonesia’s biodiversity policy. Twitter data focusing on topics such as food security, health, and environmental management were collected, with a representative subset of 13,435 tweets annotated from a larger dataset of 500,000 to ensure reliable sentiment labels through majority voting. IndoBERTweet was compared to seven traditional machine-learning classifiers using TF-IDF and BERT embeddings for feature extraction. Model performance was assessed using mean accuracy, mean F1 score, and statistical significance (p-values). Additionally, sentiment analysis included word attribution techniques with BERT embeddings, enhancing relevance, interpretability, and consistent attribution to deliver accurate insights. IndoBERTweet models consistently outperformed traditional methods in both accuracy and F1 score. While BERT embeddings boosted performance for conventional models, IndoBERTweet delivered superior results, with p-values below 0.05 confirming statistical significance. This approach demonstrates that the model’s outputs are explainable and align with human understanding. Findings underscore IndoBERTweet’s substantial impact on advancing sentiment analysis technology, showcasing its potential to drive innovation and elevate practices in the field.
- 1 June 2025
- Published by Institute of Advanced Engineering and Science in IAES International Journal of Artificial Intelligence (IJ-AI)
- Vol. 14 (3) , 2389-2401
Open AccessModeling sentiment analysis of Indonesian biodiversity policy Tweets using IndoBERTweet
This study develops and evaluates a sentiment analysis model using IndoBERTweet to analyze Twitter data on Indonesia’s biodiversity policy. Twitter data focusing on topics such as food security, health, and environmental management were collected, with a representative subset of 13,435 tweets annotated from a larger dataset of 500,000 to ensure reliable sentiment labels through majority voting. IndoBERTweet was compared to seven traditional machine-learning classifiers using TF-IDF and BERT embeddings for feature extraction. Model performance was assessed using mean accuracy, mean F1 score, and statistical significance (p-values). Additionally, sentiment analysis included word attribution techniques with BERT embeddings, enhancing relevance, interpretability, and consistent attribution to deliver accurate insights. IndoBERTweet models consistently outperformed traditional methods in both accuracy and F1 score. While BERT embeddings boosted performance for conventional models, IndoBERTweet delivered superior results, with p-values below 0.05 confirming statistical significance. This approach demonstrates that the model’s outputs are explainable and align with human understanding. Findings underscore IndoBERTweet’s substantial impact on advancing sentiment analysis technology, showcasing its potential to drive innovation and elevate practices in the field.
- Open Access
- 22 March 2025
- Published by Universitas Kristen Satya Wacana in AITI
- Vol. 22 (1) , 84-100
Pengembangan model akustik dengan deep neural network untuk sistem pengenalan wicara bahasa Indonesia
Pendekatan berbasis Deep Neural Network (DNN) menawarkan akurasi yang lebih tinggi dibandingkan dengan metode tradisional seperti Hidden Markov Model (HMM)-Gaussian Mixture Model (GMM) dalam pengembangan model akustik. Dalam penelitian ini, tiga varian DNN populer dievaluasi: Time-Delay Neural Network (TDNN), Long Short-Term Memory (LSTM), dan kombinasi hibrid TDNN-LSTM untuk pengembangan model akustik dalam pengenalan wicara bahasa Indonesia. Data wicara yang digunakan adalah KDW-BPPT-50K-ASR1 dengan durasi lebih dari 92 jam, model akustik dilatih dan eksperimen dilakukan untuk menganalisis kinerjanya. Hasil penelitian menunjukkan bahwa model hibrid TDNN-LSTM mencapai kinerja terbaik dengan Word Error Rate (WER) sebesar 9,67%, mengungguli TDNN dengan WER 12,16% dan LSTM dengan WER 10,6%. Penemuan ini menegaskan bahwa model hibrid mampu meningkatkan akurasi pengenalan wicara bahasa Indonesia dibandingkan dengan penggunaan TDNN atau LSTM secara terpisah. Hasil ini memberikan kontribusi penting bagi pengembangan sistem pengenalan wicara yang lebih akurat dan efisien.
- 22 March 2025
- Published by Universitas Kristen Satya Wacana in AITI
- Vol. 22 (1) , 84-100
Open AccessPengembangan model akustik dengan deep neural network untuk sistem pengenalan wicara bahasa Indonesia
Pendekatan berbasis Deep Neural Network (DNN) menawarkan akurasi yang lebih tinggi dibandingkan dengan metode tradisional seperti Hidden Markov Model (HMM)-Gaussian Mixture Model (GMM) dalam pengembangan model akustik. Dalam penelitian ini, tiga varian DNN populer dievaluasi: Time-Delay Neural Network (TDNN), Long Short-Term Memory (LSTM), dan kombinasi hibrid TDNN-LSTM untuk pengembangan model akustik dalam pengenalan wicara bahasa Indonesia. Data wicara yang digunakan adalah KDW-BPPT-50K-ASR1 dengan durasi lebih dari 92 jam, model akustik dilatih dan eksperimen dilakukan untuk menganalisis kinerjanya. Hasil penelitian menunjukkan bahwa model hibrid TDNN-LSTM mencapai kinerja terbaik dengan Word Error Rate (WER) sebesar 9,67%, mengungguli TDNN dengan WER 12,16% dan LSTM dengan WER 10,6%. Penemuan ini menegaskan bahwa model hibrid mampu meningkatkan akurasi pengenalan wicara bahasa Indonesia dibandingkan dengan penggunaan TDNN atau LSTM secara terpisah. Hasil ini memberikan kontribusi penting bagi pengembangan sistem pengenalan wicara yang lebih akurat dan efisien.
- 23 December 2024
- Published by IGI Global Scientific Publishing
- p. 219-262
Artificial Intelligence Technologies in Mental Health
This chapter explores the incorporation of artificial intelligence (AI) into mental health care, with a particular focus on managing depression. AI has significantly enhanced the promotion, detection, diagnosis, treatment, and monitoring of depression by leveraging technologies such as machine learning, natural language processing, and wearable devices. This chapter also discusses various AI-driven approaches, including the analysis of questionnaires, medical records, social media, speech data, electroencephalogram, magnetic resonance imaging, chatbots, virtual reality, face analysis, robots, multimodal methods, and wearable devices. Each of these technologies offers unique benefits, such as increased accuracy in detecting depression, personalized treatment plans, and continuous patient monitoring. However, the challenges linked to AI in mental health, such as data privacy issues, biases in algorithms, and the complexity of human emotions. The chapter concludes by highlighting the opportunities and future research directions and innovation for AI in enhancing of depression care.
- 23 December 2024
- Published by IGI Global Scientific Publishing
- p. 219-262
Artificial Intelligence Technologies in Mental Health
This chapter explores the incorporation of artificial intelligence (AI) into mental health care, with a particular focus on managing depression. AI has significantly enhanced the promotion, detection, diagnosis, treatment, and monitoring of depression by leveraging technologies such as machine learning, natural language processing, and wearable devices. This chapter also discusses various AI-driven approaches, including the analysis of questionnaires, medical records, social media, speech data, electroencephalogram, magnetic resonance imaging, chatbots, virtual reality, face analysis, robots, multimodal methods, and wearable devices. Each of these technologies offers unique benefits, such as increased accuracy in detecting depression, personalized treatment plans, and continuous patient monitoring. However, the challenges linked to AI in mental health, such as data privacy issues, biases in algorithms, and the complexity of human emotions. The chapter concludes by highlighting the opportunities and future research directions and innovation for AI in enhancing of depression care.
- Open Access
- 22 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 204-210
Sentiment Analysis of Indonesian New Capitol (IKN) Tweets by Stacked Generalization of Deep Learning
The increasing use of Twitter for conveying the general public's sentiment toward a specific public policy generates pros and cons and has led to much research in sentiment analysis. Instead of exploring the most suitable classifier for a sentiment analysis model individually, there is a trend of employing an ensemble of classifiers to improve the accuracy and performance of the model. We proposed a model, initially by training word embedding using word2vec from 12.5K Indonesian Twitter on the relocation issue of the new capitol city of Indonesia (IKN) and by utilizing CNN, Bidirectional LSTM, and MLP as the base classifiers. Finally, we performed a stack generalization ensemble technique using MLP and LR as the meta-classifiers and compared the performance of the ensemble techniques with individual base classifiers. The base classifiers take advantage of the weights the word embedding provides to do the learning process. The results show that the stacking ensemble using MLP performs slightly better than LR as the meta-classifier, with the F-1 score of 74.65% vs. 73.78%, respectively. MLP meta-classifiers also perform somewhat better than the hard and soft majority voting ensemble with difference F-1 scores of 3.75% and 2.56%, respectively. The results show that the proposed stacked generalization technique model has improved the performance of the sentiment analysis model.
- 22 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 204-210
Open AccessSentiment Analysis of Indonesian New Capitol (IKN) Tweets by Stacked Generalization of Deep Learning
The increasing use of Twitter for conveying the general public's sentiment toward a specific public policy generates pros and cons and has led to much research in sentiment analysis. Instead of exploring the most suitable classifier for a sentiment analysis model individually, there is a trend of employing an ensemble of classifiers to improve the accuracy and performance of the model. We proposed a model, initially by training word embedding using word2vec from 12.5K Indonesian Twitter on the relocation issue of the new capitol city of Indonesia (IKN) and by utilizing CNN, Bidirectional LSTM, and MLP as the base classifiers. Finally, we performed a stack generalization ensemble technique using MLP and LR as the meta-classifiers and compared the performance of the ensemble techniques with individual base classifiers. The base classifiers take advantage of the weights the word embedding provides to do the learning process. The results show that the stacking ensemble using MLP performs slightly better than LR as the meta-classifier, with the F-1 score of 74.65% vs. 73.78%, respectively. MLP meta-classifiers also perform somewhat better than the hard and soft majority voting ensemble with difference F-1 scores of 3.75% and 2.56%, respectively. The results show that the proposed stacked generalization technique model has improved the performance of the sentiment analysis model.
- Open Access
- 22 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 229-234
Developing Semantic Annotation Representation of Social Media Sentiments and Metadata as Resource Description Framework
Social Media has become a tool abiding the press in this modern society. Everyone can write their minds and build their mass media to publish opinions. Thus, in this manuscript, we develop a resource description framework scheme (RDFS) to enrich the information and metadata from Indonesian tweets regarding their New Capitol. This work focused on applying a popular method (i.e., the Tweetskb scheme) to construct the RDF of those tweets. We also developed the Schema to fulfill our need to contain all the information to RDF. RDF Triples were generated by connecting several established vocabularies to ensure the connection between its related nodes has meaning. The sentiment polarity (i.e., neutral, positive, and negative sentiment) is used in this manuscript. Thus, our proposal can be used as an initial work to make use of twitter's metadata to predict how reliable a user is, how the community interact with a certain topic, spam detection, clustering, and even implementing machine learning and deep learning sentiment analysis in a manner of knowledge graph.
- 22 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 229-234
Open AccessDeveloping Semantic Annotation Representation of Social Media Sentiments and Metadata as Resource Description Framework
Social Media has become a tool abiding the press in this modern society. Everyone can write their minds and build their mass media to publish opinions. Thus, in this manuscript, we develop a resource description framework scheme (RDFS) to enrich the information and metadata from Indonesian tweets regarding their New Capitol. This work focused on applying a popular method (i.e., the Tweetskb scheme) to construct the RDF of those tweets. We also developed the Schema to fulfill our need to contain all the information to RDF. RDF Triples were generated by connecting several established vocabularies to ensure the connection between its related nodes has meaning. The sentiment polarity (i.e., neutral, positive, and negative sentiment) is used in this manuscript. Thus, our proposal can be used as an initial work to make use of twitter's metadata to predict how reliable a user is, how the community interact with a certain topic, spam detection, clustering, and even implementing machine learning and deep learning sentiment analysis in a manner of knowledge graph.