agiuz‑stt‑v4‑ct2
The most accurate Uzbek Speech‑to‑Text model. Whisper‑based, converted to CTranslate2 (faster‑whisper) for fast inference.
⚠️ The model weights are proprietary (not open‑source). This page documents the model — you can try it live, but the weights are not downloadable.
🔊 Live demo · Demo · Jonli demo
👉 https://huggingface.co/spaces/sardorb3k/agiuz-stt-v4 — interface in English / Русский / Oʻzbek
English
agiuz‑stt‑v4‑ct2 transcribes Uzbek speech to text with high accuracy. It is fine‑tuned for Uzbek and tuned for real‑world audio (lectures, meetings, technical talk).
Highlights
- 🎯 High accuracy on Uzbek (Latin script)
- ✍️ Correct apostrophes —
oʻ,gʻ,maʼlumot,taʼlim - 🔤 English tech terms kept verbatim —
PDF,API,GitHub,Python,Wi‑Fi, … - 🧠 Domain decoder prompt (university, IT, programming)
- 🧹 Automatic hallucination / prompt‑echo filtering
Use it (via the hosted demo / API)
from gradio_client import Client, handle_file
client = Client("sardorb3k/agiuz-stt-v4")
text, info = client.predict(handle_file("audio.wav"), "uz", api_name="/predict")
print(text)
| Architecture | Whisper (fine‑tuned) |
| Format | CTranslate2 / faster‑whisper |
| Quantization | int8 (CPU) / float16 (GPU) |
| Language | Uzbek (uz) |
| Task | transcribe |
Русский
agiuz‑stt‑v4‑ct2 — модель распознавания узбекской речи с высокой точностью. Дообучена для узбекского языка и настроена под реальное аудио (лекции, встречи, технические беседы).
Возможности
- 🎯 Высокая точность для узбекского (латиница)
- ✍️ Корректные апострофы —
oʻ,gʻ,maʼlumot - 🔤 Английские технические термины в оригинале —
PDF,API,GitHub, … - 🧠 Доменный декодер‑промпт (университет, IT, программирование)
- 🧹 Автоматическая фильтрация галлюцинаций
Использование (через размещённое демо / API)
from gradio_client import Client, handle_file
client = Client("sardorb3k/agiuz-stt-v4")
text, info = client.predict(handle_file("audio.wav"), "uz", api_name="/predict")
print(text)
Oʻzbek
agiuz‑stt‑v4‑ct2 — oʻzbek nutqini yuqori aniqlikda matnga oʻgiruvchi model. Oʻzbek tili uchun maxsus oʻqitilgan va real audioga (maʼruza, uchrashuv, texnik suhbat) moslangan.
Imkoniyatlari
- 🎯 Oʻzbek tili uchun yuqori aniqlik (lotin yozuvi)
- ✍️ Toʻgʻri apostrof —
oʻ,gʻ,maʼlumot,taʼlim - 🔤 Inglizcha texnik atamalar asl koʻrinishda —
PDF,API,GitHub, … - 🧠 Domen decoder prompti (universitet, IT, dasturlash)
- 🧹 Gallyutsinatsiya / prompt‑echo segmentlarini avtomatik filtrlash
Foydalanish (jonli demo / API orqali)
from gradio_client import Client, handle_file
client = Client("sardorb3k/agiuz-stt-v4")
text, info = client.predict(handle_file("audio.wav"), "uz", api_name="/predict")
print(text)
👤 Author / Автор / Muallif · 📬 Contact
agiuz — @sardorb3k ✉️ s.sardorb3k@gmail.com — for questions, licensing or collaboration / по вопросам и лицензированию / savol, litsenziya yoki hamkorlik uchun
📄 License / Лицензия / Litsenziya
Proprietary — all rights reserved. See LICENSE.
Weights are not distributed; commercial or other use requires written permission —
contact s.sardorb3k@gmail.com.