Post by NavAI

🚀 NavAI jamoasi taqdim etadi! Uzbek: Biz Whisper arxitekturasi asosidagi Uzbek ASR modellar oilasini ommaga ochiqladik. Nutqni matnga aylantirish (ASR) sohasida o'zbek tili uchun katta bo'shliq bor edi - buni to'ldirishga qaror qildik. Muammo shunda ediki, Huggingface'da 100+ Whisper-asosli o'zbekcha model bor, lekin aksariyati to'liq o'qitilmagan yoki ularni o'qitilishi haqida ma'lumotlar ochiq emas. Yechimimiz: - 100% ochiq ma'lumotlar bazasida o'qitlgan model, faqat inson tomonidan belgilangan ma'lumotlar - Pseudo-label yo'q, mashinga tomonidan generatsiya qilingan transkript yo'q 🎯 Maqsad: faqat ochiq ma'lumotlar bilan qanchalik uzoqqa borish mumkinligini isbotlash Natijada: 🏆 Benchmarklarda TOP-3 (Whisper modellar oilasi bo'yicha) 📦 4 ta o'lchamlik modellar: tiny, base, small, medium 📈 Parametr oshgan sari, sifat ham izchil oshadi Har kim o'z qurilma (hardware) imkoniyati va tezlik/aniqlik ehtiyojiga qarab mos modelni tanlab, sinab ko'rishi yoki qo'shimcha o'rgatish (finetune) qilishi mumkin. 👇 Modellar, kod va to'liq benchmark natijalari - ilovada https://lnkd.in/eMXF7b5k ------------------------------------------------------------------------------ English: 🚀 NavAI introduces: Whisper-based Uzbek ASR model family There's been a significant gap in Uzbek-language speech-to-text (ASR) — we set out to close it. The problem: Huggingface hosts 100+ Whisper-based Uzbek models, but most are undertrained or has little information about the training data, and not proper benchmarking. Our approach: - we used 100% publicly available, human-labeled data - we excluded pseudo-labels, machine-generated transcripts 🎯 Goal: prove how far open data alone can take Uzbek ASR — with clean licensing and transparent benchmarks The results: 🏆 TOP-3 ranking among Whisper-based models on public benchmarks 📦 4 model sizes released: tiny, base, small, medium 📈 Performance scales consistently with model size Choose the model that fits your hardware and speed/accuracy needs — test it, or fine-tune it further for your own use case. 👇 Models, code, and full benchmark details in the link below https://lnkd.in/eMXF7b5k #UzbekAI #SpeechRecognition #ASR #OpenSource #Whisper #NLP #AIforGood