JURTEKSI (Jurnal Teknologi dan Sistem Informas.
ISSN 2407-1811 (Prin.
Vol.
XII No 3.
June 2026, hlm.
429 Ae 436 ISSN 2550-0201 (Onlin.
DOI: http://dx.
org/ 10.
33330/jurteksi.
Available online at http://jurnal.
id/index.
php/jurteksi PERFORMANCE EVALUATION OF AUTOMATED MEETING
SUMMARIZATION BASED ON OPENAI WHISPER
AND INDOT5 FINE-TUNING
I Gusti Lanang Oka Wiyana1* .
Putu Indah Ciptayani 1 .
Ida Bagus Adisimakrisna Peling 1 Teknologi Rekayasa Perangkat Lunak.
Politeknik Negeri Bali email: *gungwiyana@gmail.
Abstract: Manual meeting documentation risks losing important information due to cognitive Although automated summarization models have evolved, integrated end-to-end systems for Indonesian spoken language remain highly limited.
This study aims to design and evaluate an end-to-end automated meeting summarization architecture that directly integrates Automatic Speech Recognition (ASR) via OpenAI Whisper for transcription and the IndoT5 language model for abstractive summarization.
IndoT5 was fine-tuned using a dataset of 486 Indonesian spoken language transcript pairs.
Testing was conducted on a CPU infrastructure using MP4.
MP3, and WAV formats.
Results show the optimal fine-tuning configuration significantly improved accuracy, achieving ROUGE-1 .
ROUGE-2 .
, and ROUGE-L .
Computationally, the system achieved a Real-Time Factor below 1, processing data faster than the actual recording duration.
Conclusively, integrating Whisper and IndoT5 shows potential in producing coherent meeting summaries with lightweight computational overhead, making it viable for local infrastructure implementation to ensure data privacy.
Keywords: abstractive summarization.
ASR.
end-to-end pipeline.
IndoT5.
real-time factor Abstrak: Dokumentasi rapat manual rentan menghilangkan informasi penting akibat keterbatasan kognitif.
Meskipun model peringkas otomatis telah berkembang, implementasi sistem terintegrasi .
nd-to-en.
khusus percakapan lisan berbahasa Indonesia masih sangat Penelitian ini bertujuan merancang dan mengevaluasi arsitektur peringkas rapat otomatis end-to-end yang mengintegrasikan langsung Automatic Speech Recognition (ASR) melalui OpenAI Whisper untuk transkripsi dan model bahasa IndoT5 untuk peringkasan Adaptasi domain dilakukan melalui fine-tuning IndoT5 menggunakan 486 pasang dataset transkrip lisan berbahasa Indonesia.
Pengujian pada infrastruktur CPU menggunakan format MP4.
MP3, dan WAV.
Hasil pengujian menunjukkan konfigurasi fine-tuning optimal berhasil meningkatkan akurasi, dengan skor ROUGE-1 .
,4.
ROUGE-2 .
,1.
, dan ROUGE-L .
,2.
Sistem mendemonstrasikan efisiensi komputasi dengan nilai Real-Time Factor di bawah 1, mengindikasikan waktu pemrosesan lebih cepat dari durasi rekaman asli.
Kesimpulannya, integrasi Whisper dan IndoT5 menunjukkan potensi dalam menghasilkan ringkasan yang koheren dengan beban komputasi ringan, sehingga layak diimplementasikan pada infrastruktur lokal organisasi untuk menjaga privasi data.
Kata kunci: ASR.
end-to-end pipeline.
IndoT5.
peringkasan abstraktif.
real-time factor
INTRODUCTION
JURTEKSI (Jurnal Teknologi dan Sistem Informas.
ISSN 2407-1811 (Prin.
Vol.
XII No 3.
June 2026, hlm.
429 Ae 436 ISSN 2550-0201 (Onlin.
DOI: http://dx.
org/ 10.
33330/jurteksi.
Available online at http://jurnal.
id/index.
php/jurteksi Meeting activities in modern organizations show a 12.
9% increase in virtual meeting frequency per person, along with a 13.
5% surge in the number of meeting participants, driven by the transition to post-pandemic work styles .
This increased utilization of video conferencing demands more efficient management of post-event deliverables.
The importance of this documentation is evident from the high organizational need to archive meeting minutes as a reference for strategic decision-making and proof .
However, the manual minuting process remains a primary constraint due to human cognitive limitations in capturing important points in real-time amidst speakers' talking speeds and lengthy discussion durations, which ultimately increases the risk of errors and the loss of critical information .
As a solution, developing an system based on the integration of Automatic Speech Recognition (ASR) and Natural Language Processing (NLP) is essential.
In the NLP domain, abstractive approaches, such as those implemented through the IndoT5 deep learning model, are considered superior compared to multilingual summarization models like mBART or conventional extractive approaches .
, .
, as they are capable of constructing new sentences that retain the core meaning with a more natural linguistic structure.
A state-ofthe-art transformer model that does not rely on sequential hidden states is required to ensure the coherence of the summary .
However, abstractive summarization models generally require pure text input, making ASR technology necessary to map sound waves into text sequences .
The OpenAI Whisper model has proven to be robust as an ASR because it is trained using massive weakly supervised methods, allowing it to accurately predict text from various speaker accents and background noise conditions without requiring rigid data standardization .
, .
Several previous studies have explored the use of pre-trained T5 family however, their effectiveness on non-standard spoken structures has not been widely tested .
Other research has attempted to apply ASR technology for audio transcription but stopped at the text extraction stage without further summarization processes .
Furthermore, transcript datasets are often implemented specifically in English, which has characteristics from Indonesian .
, .
Based on this literature review, the integration of the OpenAI Whisper ASR model and IndoT5 within an End-to-End pipeline architecture for Indonesian underexplored and has not been comprehensively evaluated.
Therefore, this study aims to design and evaluate the performance of this architecture by measuring computational time .
nference tim.
and validating text feasibility using the ROUGE metric .
The final results of this study are expected to serve as an automation solution that supports postmeeting knowledge management in JURTEKSI (Jurnal Teknologi dan Sistem Informas.
ISSN 2407-1811 (Prin.
Vol.
XII No 3.
June 2026, hlm.
429 Ae 436 ISSN 2550-0201 (Onlin.
DOI: http://dx.
org/ 10.
33330/jurteksi.
Available online at http://jurnal.
id/index.
php/jurteksi METHOD using FFmpeg.
After the data was prepared, the research proceeded to the system architecture design stage, model training experiments .
ine-tunin.
, and the final stage comprising system performance testing and evaluation analysis before drawing a conclusion.
More specifically, in the system design stage, the constructed architecture integrates two primary artificial intelligence modules executed entirely in a local environment.
The first module is speech recognition using the Medium variant of the OpenAI Whisper ASR model to achieve a balance of high transcription accuracy while avoiding network latency issues.
The text output from this module is automatically forwarded as input to the second module, an abstractive text summarizer using the pre-trained IndoT5-base model from the Wikidepia repository, which has undergone a domain adaptation process .
The technical data flow within this integrated system from the uploading of audio documents by the user to the presentation of the final summary through a Streamlitbased web interface is thoroughly illustrated in Image 2.
This study employs a quantitative experimental approach to develop and evaluate the End-to-End (E2E) pipeline architecture of an automated meeting The computational process was built using the Python 3.
10 programming language and the PyTorch deep learning framework .
, .
Generally, the research process was carried out systematically following the stages illustrated in Image Image 1.
Research Flowchart The workflow began with the collection of 30 video recordings of Plenary Meetings of the House of Representatives of the Republic of Indonesia (DPR RI).
The data was processed through chunking to extract 486 text pairs with an 80:20 split ratio to prevent data leakage, and extracted into MP3 and WAV formats Image 2.
End-to-End Pipeline Architecture of the Meeting Summarization System JURTEKSI (Jurnal Teknologi dan Sistem Informas.
ISSN 2407-1811 (Prin.
Vol.
XII No 3.
June 2026, hlm.
429 Ae 436 ISSN 2550-0201 (Onlin.
DOI: http://dx.
org/ 10.
33330/jurteksi.
Available online at http://jurnal.
id/index.
php/jurteksi The system testing stage focused on computational efficiency and linguistic quality.
Efficiency was measured through inference latency, which is the total execution time from the upload process to the output presentation.
Summary quality was evaluated by comparing the system's output text against reference documents .
round trut.
using the ROUGE (Recall-Oriented Understudy for Gisting Evaluatio.
The systematic calculation of the ROUGE metric is formulated in Equation .
Oc Oc stage is essential to ensure that the endto-end testing utilizes the most optimal The domain adaptation process utilized a secondary dataset pairs of spoken transcriptions and manual summaries in Indonesian, with a total file size of 1.
MB.
For the training process, the dataset was partitioned using an 80:20 ratio, where 80% .
data pair.
was allocated as training data and 20% .
data pair.
as testing data.
The data separation was programmatically using the train_test_split module from the scikitlearn library in Python.
It was grouped based on the original file identity so that segments from the same recording were not distributed into two different sets, aiming to prevent data leakage.
This file identity-based processing literature to ensure that model validation is free from repetitive context bias .
The comparison results of the presented in Table 1.
The data in Table 1 shows that the EXP-013 configuration serves as the baseline for the system architecture as it produced the highest linguistic accuracy.
This optimized model was consistently used in the system performance testing stability and avoid evaluation bias, inference latency was executed on 11 test data samples of varying durations.
The computational scenarios were tested on a CPU infrastructure with MP4.
MP3, and WAV input formats.
The detailed test results are presented in Table 2.
The description of Equation .
as the length of the n-gram used in the evaluation, as the n-th n-gram in the summary text, as the human reference summary, the number of matching n-grams in the system output, and as the number of n-grams in the reference.
A higher ROUGE percentage value directly represents a higher level of accuracy and a more representative informational similarity between the computational system and the reference .
RESULT AND DISCUSSION
The implementation of artificial intelligence in this system began with determining the model configuration fine-tuning experiments on the IndoT5 model.
This JURTEKSI (Jurnal Teknologi dan Sistem Informas.
ISSN 2407-1811 (Prin.
Vol.
XII No 3.
June 2026, hlm.
429 Ae 436 ISSN 2550-0201 (Onlin.
DOI: http://dx.
org/ 10.
33330/jurteksi.
Available online at http://jurnal.
id/index.
php/jurteksi Table 1.
IndoT5 Model Fine-Tuning Hyperparameter Experiment Results ROUGE
Batch
Experiment ID
Epoch
Size
EXP-BASE
EXP-001
EXP-002
EXP-003
EXP-004
EXP-005
EXP-006
EXP-007
EXP-008
EXP-009
EXP-010
EXP-011
EXP-012
EXP-013
EXP-014
EXP-015
EXP-016
EXP-017
EXP-018
Table 2.
Inference Latency Testing on 11 Test Data Samples Actual Duration MP4 Time MP3 Time WAV Time Data Sample .
Data 1 Data 2 Data 3 Data 4 Data 5 Data 6 Data 7 Data 8 Data 9 Data 10 Data 11 Average The data in Table 2 demonstrates the system's performance stability across various recording durations.
On average, computing the MP4 format required 42 seconds, and the MP3 format 65 seconds.
Meanwhile, the WAV format recorded the most efficient seconds.
All format variations achieved a Real-Time Factor index below 1, indicating that the system is capable of processing the minute transcripts faster than the original audio duration.
The achievement of this computational time considering previous studies emphasized that the primary obstacle in implementing artificial intelligence models in realworld scenarios is often constrained by high computational loads and processing latency .
The efficiency in the WAV format occurs because the acoustic data is lossless, allowing the model to extract spectral representations without the bottleneck of heavy audio decompression In addition to time analysis, the reconstruction quality was measured against 11 reference document samples ROUGE The performance comparison of average linguistic accuracy across all test data is presented in Table 3.
MP3 files, which reduces certain frequency spectrum ranges to minimize memory size.
This condition produces microscopic acoustic distortions that cause the model to predict phonemes with slight deviations.
The accuracy degradation due to signal compression aligns with previous research showing that maintaining audio signal integrity .
is crucial for preserving feature representation quality, thereby improving Transformer-based model performance in the information extraction process .
The stable average ROUGE-2 acquisition at around 0.
20 reflects the characteristics of abstractive summarization, where the IndoT5 model performs high-level semantic reconstruction by paraphrasing the main ideas using new lexical structures rather than merely performing This abstractive capability is highly relevant for reducing narrative redundancy in unstructured spoken conversations into professional minute documents.
Table 3.
Average ROUGE Metric
Evaluation Across Formats ROUGE
Format
MP4
MP3
WAV
CONCLUSION
The integration of the OpenAI Whisper Medium variant model and the IndoT5 language model, which has undergone a domain adaptation process.
Based on the evaluation results, the fine-tuning process was proven to significantly improve linguistic quality compared to the baseline model, with an increase in ROUGE-1 from 2484 to 0.
ROUGE-2 from 0.
1973, and ROUGE-L from 0.
1637 to The demonstrated by an average Real-Time Factor value below one, indicates that the system has the potential to be The evaluation results in Table 3 validate that the input representations from the MP4 and WAV formats identical unigram similarity level of On the other hand, the compressed MP3 format experienced a slight decrease in the ROUGE-1 score to This phenomenon is closely related to the lossy compression nature of implemented processing speed and data privacy.
As a future development effort, the integration of a Speaker Diarization feature can be enhance the system's functionality in automatically identifying speaker transitions in multi-participant meeting scenarios to produce more informative meeting minutes.
BIBLIOGRAPHY