Lyla Ruslana Aini
Research Subjects
Lyla Ruslana Aini has a total of 15 co-authors. Here are the top co-authors by number of shared publications:
- Agung Santosa
National Research and Innovation Agency (BRIN)
3 shared publications - Mohammad Teduh Uliniansyah
National Research and Innovation Agency (BRIN)
3 shared publications - Elvira Nurfadhilah
Brin, Indonesia
2 shared publications - Gunarso Gunarso
National Research and Innovation Agency (BRIN)
2 shared publications - Siska Pebiana
National Research and Innovation Agency, Indonesia
2 shared publications - Ahmad Fudholi
National Research and Innovation Agency
1 shared publication
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.
- 1 April 2026
- Published by Ital Publication in Emerging Science Journal
- Vol. 10 (2) , 866-892
Multimodal Emotion Recognition Using Hybrid Large Language Models and Metaheuristic Algorithms
Emotion recognition is a vital component of human–computer interaction and intelligent systems, yet robust multimodal emotion recognition remains challenging due to high-dimensional input space, noisy features, and the complexity of integrating heterogeneous modalities. This study proposes a novel hybrid multimodal framework that enhances both accuracy and computational efficiency by combining the semantic representation capability of Large Language Models (LLMs) with the optimization strengths of metaheuristic algorithms. In the proposed approach, an LLM is utilized to extract high-level contextual features from text and audio streams, while the Binary Artificial Hummingbird Algorithm (BAHA) performs feature selection to remove redundant attributes. Subsequently, the Goose Algorithm (GA) optimizes classifier hyperparameters, and the Komodo Mlipir Algorithm (KMA) conducts late fusion of the final multimodal outputs. Experiments conducted on the Interactive Emotional Dyadic Motion Capture (IEMOCAP) dataset, evaluated on six emotion categories, demonstrate that this hybrid approach successfully captures subtle affective cues and surpasses state-of-the-art baselines, achieving an accuracy of 87.5%. Integrating LLMs with multiple specialized metaheuristics therefore yields a substantially more robust emotion recognition pipeline and represents a promising direction toward the development of more emotionally intelligent systems.
- 1 April 2026
- Published by Ital Publication in Emerging Science Journal
- Vol. 10 (2) , 866-892
Multimodal Emotion Recognition Using Hybrid Large Language Models and Metaheuristic Algorithms
Emotion recognition is a vital component of human–computer interaction and intelligent systems, yet robust multimodal emotion recognition remains challenging due to high-dimensional input space, noisy features, and the complexity of integrating heterogeneous modalities. This study proposes a novel hybrid multimodal framework that enhances both accuracy and computational efficiency by combining the semantic representation capability of Large Language Models (LLMs) with the optimization strengths of metaheuristic algorithms. In the proposed approach, an LLM is utilized to extract high-level contextual features from text and audio streams, while the Binary Artificial Hummingbird Algorithm (BAHA) performs feature selection to remove redundant attributes. Subsequently, the Goose Algorithm (GA) optimizes classifier hyperparameters, and the Komodo Mlipir Algorithm (KMA) conducts late fusion of the final multimodal outputs. Experiments conducted on the Interactive Emotional Dyadic Motion Capture (IEMOCAP) dataset, evaluated on six emotion categories, demonstrate that this hybrid approach successfully captures subtle affective cues and surpasses state-of-the-art baselines, achieving an accuracy of 87.5%. Integrating LLMs with multiple specialized metaheuristics therefore yields a substantially more robust emotion recognition pipeline and represents a promising direction toward the development of more emotionally intelligent systems.