Electronic Theses and Dissertation

Kelancaran membaca (reading fluency) siswa indonesia masih memprihatinkan, tercermin dari peringkat ke-71 dari 81 negara pada skor pisa 2022. penilaian kelancaran membaca manual oleh guru bersifat tidak efisien dan tidak skalabel, sementara model automatic speech recognition (asr) seperti whisper belum dioptimasi untuk suara penutur muda bahasa indonesia. penelitian ini bertujuan menganalisis performa model whisper pretrained dalam mentranskripsikan suara siswa sd dan smp, mengembangkan model tersebut melalui fine-tuning dan optimasi whisper.cpp agar efisien pada perangkat mobile, serta mengembangkan dan menguji dampak aplikasi game pelatihan kelancaran membaca bernama readrace. model whisper tiny di-fine-tuning menggunakan 16.588 sampel suara penutur muda dari korpus mozilla common voice versi 25.0 menggunakan speaker-aware splitting untuk mencegah kebocoran data antar subset, lalu dikonversi dan dikuantisasi menggunakan whisper.cpp. hasil penelitian menunjukkan model pretrained menghasilkan word error rate (wer) 63,98% dan character error rate (cer) 41,86%, yang menurun signifikan setelah fine-tuning menjadi wer 21,67% dan cer 7,60%, dengan varian kuantisasi q8_0 (41,52 mb) menawarkan keseimbangan terbaik antara ukuran dan akurasi dibandingkan model asli (±151 mb). aplikasi readrace berhasil mengintegrasikan model tersebut untuk transkripsi suara secara on-device, dengan seluruh 31 skenario blackbox testing dinyatakan berhasil (pass). pengujian dampak aplikasi dengan menggunakan metode pretest-posttest control group memiliki pengaruh signifikan dari hasil uji-t sampel bebas terhadap peningkatan kelancaran membaca siswa. penelitian ini menunjukkan bahwa kombinasi fine-tuning dan optimasi whisper.cpp menghasilkan model asr bahasa indonesia yang akurat dan ringan, bermanfaat sebagai dasar pengembangan alat evaluasi kelancaran membaca yang objektif, efisien, dan dapat diakses sekolah tanpa koneksi internet. kata kunci: automatic speech recognition, whisper, fine-tuning, whisper.cpp, kelancaran membaca, pretest-posttest control group.

Reading fluency among Indonesian students remains concerning, reflected by Indonesia's 71st-of-81 ranking in the 2022 PISA reading score. Conventional manual teacher-based reading fluency assessment is inefficient and hard to scale, while existing Automatic Speech Recognition (ASR) models such as Whisper remain unoptimized for young Indonesian speakers. This study analyzes pretrained Whisper's performance in transcribing elementary and junior high school students' speech, develops the model through fine-tuning and whisper.cpp-based optimization for mobile deployment, and develops and tests the impact of a mobile game-based reading fluency training application named ReadRace. The Whisper Tiny model was fine-tuned on 16,588 young Indonesian speaker speech samples from Mozilla Common Voice v25.0, using speaker-aware splitting to prevent data leakage across subsets. The pretrained model produced a Word Error Rate (WER) of 63.98% and Character Error Rate (CER) of 41.86%, decreasing significantly after fine-tuning to 21.67% WER and 7.60% CER. The model was converted and quantized via whisper.cpp into five variants sized 24.15-41.52 MB, far smaller than the original ±151 MB model, with the q8_0 variant offering the best size-accuracy trade-off. The ReadRace application integrated the optimized model for on-device transcription, and all 31 blackbox testing scenarios were reported as Pass. Impact testing using a pretest-posttest control group design showed a significant effect based on independent sample t-test results, indicating improvement in students' reading fluency. Overall, combining Whisper fine-tuning with whisper.cpp optimization produces a more accurate, lightweight Indonesian ASR model, providing a foundation for a more objective, efficient, and accessible offline reading fluency assessment tool for schools. Keywords: Automatic Speech Recognition, Whisper, Fine-Tuning, Whisper.cpp, Reading Fluency, Pretest-Posttest Control Group.