Jonibek21/Zehnova-Uzbek-STT

Zehnova STT — O'zbek tili uchun Speech-to-Text modeli

O'zbek tili uchun fine-tune qilingan Whisper Medium asosidagi avtomatik nutqni matnга aylantirish modeli.

Model haqida

• Model turi: Automatic Speech Recognition (ASR)
• Asos model: Kotib/uzbek_stt_v1 (Whisper Medium)
• Fine-tuning usuli: LoRA (Low-Rank Adaptation)
• Til: O'zbek tili 🇺🇿
• Muallif: Jonibek21

Ishlatish

from transformers import WhisperForConditionalGeneration, WhisperProcessor, pipeline
import torch

model_id = "Jonibek21/Zehnova-stt-uzbek"

model = WhisperForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.float16
).to("cuda")

processor = WhisperProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    chunk_length_s=30,
    stride_length_s=5,
    batch_size=4,
    device=0,
)

result = pipe(
    "audio.wav",
    generate_kwargs={
        "language": "uz",
        "task": "transcribe",
        "no_repeat_ngram_size": 3
    }
)

print(result["text"])

Training ma'lumotlari

• Dataset: Maxsus O'zbek tili audio dataseti
• Train samples: 9,214
• Test samples: 1,024
• Dataset vaqti: 16 soat
• Training hardware: NVIDIA RTX 3090 (24GB)
• Training framework: Hugging Face Transformers + PEFT
• Precision: fp16
• LoRA rank: 32
• LoRA alpha: 64
• LoRA target modules: q_proj, v_proj

📊 Model Evaluation (WER)

Category	WER
Overall	~11-13%
Clean Speech	~6-11%
Noisy/Augme	~9-16%
News / Formal	~11-12%

Base model (Kotib/uzbek_stt_v1) overall WER: 16.7% Zehnova modeli base modeldan ~5% yaxshiroq natija ko'rsatdi.

Cheklovlar

• Faqat o'zbek tilida ishlaydi
• Shovqinli audio da sifat pasayishi mumkin
• 30 soniyadan uzun audiolar bo'laklarga bo'linadi

Date

• 01/05/2026