lynx9844/whisper-large-v3-turbo-uzbek-v1

🇺🇿 Whisper Large v3 Turbo Uzbek (v1)

Bu model OpenAI Whisper Large v3 Turbo arxitekturasi asosida yaratilgan bo'lib, o'zbek tili nutqini matnga aylantirish (Speech-to-Text / STT) uchun maxsus sozlangan va optimallashtirilgan 1-versiyadir (v1).

Ushbu reliz butun o'zbek ochiq kodli dasturchilar hamjamiyati (open-source community) uchun ochiq taqdim etilmoqda. Maqsad — har bir tadqiqotchi va dasturchi noldan katta hisoblash quvvati sarflamasdan, tayyor yuqori sifatli o'zbek STT modelidan foydalana olishi.

👨‍💻 Muallif & Identifikatsiya (Author & Provenance)

Kelajakdagi barcha o'zbek nutq va AI loyihalari uchun rasmiy mualliflik imzosi:

• —Muallif (Author): R. Nematov
• —Taxallus (Handle / Alias): Lynx (@lynx9844)
• —GitHub: github.com/nematov9844
• —Bog'lanish (Email): nematov9844@gmail.com
• —Yo'nalish (Initiative): Mustaqil O'zbek AI Tadqiqoti (Independent Uzbek Speech AI Initiative)

🎯 Model Haqida va Qilingan Ishlar

1. 1.Baza Model: openai/whisper-large-v3-turbo (800M parametr, OpenAI'ning eng tezkor flagmani).
2. 2.O'zbek tiliga moslashtirish (Finetuning & Conditioning):
3. 3.Universal Whisper modelining o'zbek tilidagi kamchiliklari (turkcha/qirg'izcha tovushlarga chalg'ishi) bartaraf etildi.
4. 4.O'zbek tili alifbosi, tutuq belgilari (', ʻ, ʼ), jarangli/jarangsiz tovushlar va keng tarqalgan so'zlashuv iboralari uchun maxsus yo'naltiruvchi dekoder (prompt conditioning & vocabulary alignment) integratsiya qilindi.
5. 5.Ma'lumotlar bazasi (Dataset): O'zbekistonning turli mintaqaviy lahjalari va aksentlarini o'z ichiga olgan DavronSherbaev/uzbekvoice-filtered hamda saralangan o'zbek audio to'plamlari orqali tahlil qilingan.

⚖️ Sifat Darajasi va Xolisona Baho (Neutral Evaluation)

🚀 Qanday Foydalanish Mumkin? (Quickstart)

Modelni o'z loyihalaringizga oddiy va tez ulash mumkin:

Python (Transformers kutubxonasi orqali):

import torch
import torchaudio
from transformers import WhisperProcessor, WhisperForConditionalGeneration

# 1. Model va Processorni yuklash
model_id = "lynx9844/whisper-large-v3-turbo-uzbek-v1"
device = "cuda:0" if torch.cuda.is_available() else "cpu"

processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.float16 if device != "cpu" else torch.float32,
    low_cpu_mem_usage=True
).to(device)

# 2. Audioni yuklash (16kHz mono tavsiya etiladi)
speech_tensor, rate = torchaudio.load("audio.ogg")
if rate != 16000:
    import torchaudio.transforms as T
    speech_tensor = T.Resample(rate, 16000)(speech_tensor)
if speech_tensor.shape[0] > 1:
    speech_tensor = speech_tensor.mean(dim=0, keepdim=True)

inputs = processor(speech_tensor.squeeze().numpy(), sampling_rate=16000, return_tensors="pt").input_features.to(device)
if device != "cpu":
    inputs = inputs.half()

# 3. Transkripsiya qilish
with torch.no_grad():
    predicted_ids = model.generate(
        inputs,
        language="uz",
        task="transcribe",
        no_repeat_ngram_size=3
    )

text = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print("Transkripsiya:", text)

📜 Litsenziya

Ushbu loyiha Apache-2.0 litsenziyasi ostida ochiq tarqatiladi. Siz undan tijoriy va notijoriy loyihalarda erkin foydalanishingiz mumkin.