🎙️ Vosen-STT-Uz-Eng-Rus: Robust Multilingual Speech Recognition (Low-Volume & Code-Switching Optimized)
Vosen-STT-Uz-Eng-Rus — bu past ovozda shivirlab aytilgan (low-volume, whispered) nutqlarni va tillar aralashib kelgan (code-switching) gaplarni yuqori aniqlikda transkripsiya qilishga moslashtirilgan ko'p tilli o'zbekcha Speech-to-Text modeli.
🌟 Model haqida (About the Model)
Standart Whisper modellari past va shivirlashga yaqin ovozlarda xatoliklarga ko'p yo'l qo'yadi. Ushbu muammoni bartaraf etish maqsadida, model OpenLimi tomonidan maxsus tayyorlangan ma'lumotlar to'plami yordamida qayta o'qitildi va sezilarli darajada optimallashtirildi.
Whisper me'morchiligining tabiiy ko'p tillilik xususiyati tufayli, model faqat o'zbekcha emas — o'zbek, ingliz va rus tillarida, hattoki bir gap ichida tillar aralashib kelganda ham (code-switching, masalan "keyin meeting'da discuss qilamiz", "hisobotni project bo'yicha tayyorladim") nutqni ishonchli tarzda matnga aylantira oladi.
- Ishlab chiqaruvchi: OpenLimi
- Tillar: O'zbek (uz), Ingliz (en), Rus (ru) — shu jumladan tillar aralash holatda
- Vazifa: Automatic Speech Recognition (ASR) / Speech-to-Text (STT)
- Litsenziya: MIT (Tijoriy va shaxsiy foydalanish uchun mutlaqo ochiq)
📊 O'qitish ma'lumotlari va Parametrlari (Training Data & Hyperparameters)
Ma'lumotlar to'plami
- Manba dataset:
Beehzod/uzbek_speech_data(407 ta original yuqori sifatli audio-matn juftligi) - Hajm augmentatsiyasi: Har bir audio 6 xil ovoz balandligi darajasida (+10dB, +6dB, 0dB, -6dB, -10dB, -15dB) sun'iy ravishda ko'paytirildi. Jami o'qitish bazasi: ~2,440 ta audio-matn juftligi.
- Domen: Qonunchilik, rasmiy hujjatlar uslubi va fuqarolar murojaatlari bilan bog'liq nutq.
O'qitish sozlamalari
| Parametr | Qiymat |
|---|---|
| Learning rate | 1e-5 |
| Epochs | 5 (Early stopping bilan 4-epochda to'xtatildi) |
| Batch size | 2 (Gradient accumulation 8, effektiv batch 16) |
| Optimal checkpoint | Epoch 2 |
| Mixed precision | fp16 (Half precision training) |
Baholash Natijalari (Evaluation Results)
| Epoch | Training Loss | Validation Loss | WER (Word Error Rate) |
|---|---|---|---|
| 1 | 0.1927 | 0.1058 | 11.18% |
| 2 (Tanlangan) | 0.0465 | 0.1121 | 8.82% |
| 3 | 0.0287 | 0.1185 | 10.39% |
| 4 | 0.0178 | 0.1185 | 9.97% |
Eng yaxshi natija (WER 8.82%) 2-epochda qo'lga kiritildi va overfitting boshlanishidan oldin shu checkpoint yakuniy model sifatida tanlandi.
🚀 Qanday foydalanish mumkin (Usage)
Method 1: Hugging Face Transformers orqali
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import librosa
repo_id = "openlimi/Vosen-stt-uz-eng-rus"
processor = WhisperProcessor.from_pretrained(repo_id)
model = WhisperForConditionalGeneration.from_pretrained(repo_id)
audio, sr = librosa.load("audio.wav", sr=16000)
input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
predicted_ids = model.generate(input_features, language="uz", task="transcribe")
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)
Method 2: Pipeline orqali (Tavsiya etiladi)
from transformers import pipeline
pipe = pipeline(
"automatic-speech-recognition",
model="openlimi/Vosen-stt-uz-eng-rus",
chunk_length_s=30,
device="cuda"
)
result = pipe("audio.wav", generate_kwargs={"language": "uz", "task": "transcribe"})
print(result["text"])
⚠️ Cheklovlar (Limitations)
- Domen xususiyati: Model asosan qonunchilik va rasmiy hujjat uslubidagi nutqda o'qitilgan, kundalik norasmiy suhbatlarda va o'ta texnikaviy tor doiradagi matnlarda aniqlik o'zgarishi mumkin.
- Lug'at boyligi: O'qitish datasi kichik (407 original namuna) bo'lganligi sababli, kam uchraydigan dialekt so'zlarida xatolik ehtimoli bor.
- Multilingual moslashuv: Ingliz va rus tillarini tushunish asosan Whisper asosiy arxitekturasidan keladi. Shu sababli murakkab va uzun ingliz/rus jumlalaridagi aniqlik darajasi o'zbek tilidagidan pastroq bo'lishi mumkin.
🌐 SEO Keywords & Search Engine Optimization
Uzbek STT, Uzbek ASR, Multilingual Uzbek Speech to Text, O'zbekcha inglizcha ruscha nutqni aniqlash, Code-switching Uzbek, Low-volume Whisper Uzbek, OpenLimi Vosen STT, Sun'iy intellekt matnga o'tkazish, Uzbek Whisper model, MIT Uzbek STT.