O‘ZBEK NUTQINI MATNGA AYLANTIRISH UCHUN MOSLASHTIRILGAN NUTQ TANISH MODELINI ISHLAB CHIQISH
Keywords:
avtomatik nutqni tanish, o‘zbek tili, ma’lumotlarni tozalash, mel-spektr, moslashtirib o‘qitish, isitib borish, gradientni cheklash, erta to‘xtatish, aralash aniqlik, so‘zdagi xatolik ulushiAbstract
Mazkur maqolada o‘zbek tilidagi nutqni avtomatik tarzda matnga aylantirish tizimi ishlab chiqildi, sinovdan o‘tkazildi va tahlil qilindi. Tadqiqotda ikki manba bir ovozli o‘qib eshittirish va ko‘p so‘zlovchili kundalik nutqlar birlashtirilib, matnlar me’yoriy lotin yozuviga keltirildi, audio yozuvlar 16 ming gerts namuna tezligiga moslashtirildi, uzunligi 30 soniya chegarasida boshqarildi. Ma’lumotlar mashq (119 831 yozuv), tekshiruv (648 yozuv) va sinov (899 yozuv) to‘plamlariga qat’iy ajratildi. Model ko‘p tilli yirik andozadan (Whisper-small) o‘zbek tili uchun moslashtirib o‘qitildi; isitib borish, og‘irlikni kamaytirish, gradientni cheklash va erta to‘xtatish choralaridan foydalanildi, aralash aniqlikdagi hisoblash orqali tezlik oshirildi. Baholashda so‘zdagi xatolik ulushi ko‘rsatkichi qo‘llanib, sinov to‘plamida 13,78 foiz natija olindi; sinov yo‘qotishi 0,1487, tekshiruv yo‘qotishi yakunda taxminan 0,15 bo‘ldi. Natijalar turli manbalarni birlashtirish, kiritmalarni yagona andozaga keltirish va ehtiyotkor moslashtirib o‘qitish orqali o‘zbek nutqini turli sharoitlarda barqaror tanish mumkinligini ko‘rsatadi. Ushbu yondashuv ta’lim, tibbiyot, raqamli xizmatlar va ovozli muloqot talab qiluvchi dasturlarda amaliy qo‘llash imkonini kengaytiradi.
Downloads
References
1. Radford, A. va boshq. Whisper: Robust Speech Recognition via Large-Scale Weak Supervision. OpenAI texnik hisobot , 2022.
2. Uzbek Speech Corpus va Avtomatik Nutqni Tanish. ISSAI loyiha sahifasi, 2021.
3. Uzbek Speech Corpus — dataset card. Hugging Face, 2021.
4. Povey, A.; Povey, K. FeruzaSpeech: A 60-Hour Uzbek Read Speech Corpus with Punctuation, Casing, and Context. ICNLSP 2024 materiallari, 2024.
5. Mozilla Common Voice — umumiy portal. Ma’lumotlar platformasi, 2017.
6. Common Voice Scripted Speech 23.0 — Uzbek datasheet. Mozilla Data Collective, 2024.
7. Mukhamadiyev, A. va boshq. Development of Language Models for Continuous Uzbek Speech Recognition System. Sensors (MDPI), 23(3):1145, 2023. (PMC to‘liq matn).
8. Li, J. va boshq. Improving Whisper’s Recognition Performance for Under-Represented Language (Kazakh) Leveraging Unpaired Speech and Text. Interspeech 2024 (ISCA Archive), 2024.
9. NVIDIA NeMo Framework User Guide — Checkpoints/Training results, 2024.
10. NVIDIA FastConformer-Hybrid Large (Uzbek) — model sahifasi. Hugging Face, 2024.
11. Whisper models in Amazon SageMaker JumpStart. AWS Machine Learning Blog, 10 Oct 2023.
12. Nguyen, M. T. va boshq. Improving Speech Recognition with Jargon Injection. SIGDIAL 2024 materiallari (ACL Anthology), 2024.
13. Peng, Y. va boshq. OWSM v4: Improving Open Whisper-Style Speech Models via Data Scaling and Cleaning. Interspeech 2025 (ISCA Archive), 2025.
14. Khamidov, A., Ismatov, M., Karimova, F. (2024). Integration of Whisper Models for Uzbek ASR in Noisy Environments. Journal of Modern Information Technologies and AI Research, 2(1), 58-66.
Downloads
Published
Issue
Section
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
License Terms of our Journal
