rubai-corrector-transcript-uz
Transcript-display normalization model for Uzbek ASR output with mixed Uzbek/Russian support. Built on ByT5 architecture.
This is the transcript-display variant of the rubai-corrector model family. For the fine-tuning foundation checkpoint, see rubai-corrector-base.
Authors
- Sardor Islomov — lead author
- Davron Ibrokhimov
This checkpoint is tuned for:
- display-ready punctuation and casing
- apostrophe normalization
- OCR / ASR typo cleanup
- Latin Russian -> Cyrillic Russian recovery
- mixed Uzbek/Russian transcript cleanup
- selected text-to-number normalization patterns
Intended Use
Use this model after ASR to convert noisy transcript text into better display text.
It is best for:
- Rubai-style Uzbek ASR postprocessing
- Uzbek display text cleanup
- mixed Uzbek/Russian lines where Russian appears in Latin transcription
Primary upstream ASR models this normalizer is intended to follow within the same Rubai model family:
The model is focused on line-level transcript outputs that look like the text produced by those ASR models.
It is not optimized for:
- literal no-edit transcript preservation
- noisy Gemini-style mixed-script metadata targets with forced Cyrillic inside Uzbek morphology
- aggressive general denormalization beyond the transcript-display objective
Model Family
| Model | Use Case |
|---|---|
| rubai-corrector-base | Fine-tuning base for new correction tasks |
| rubai-corrector-transcript-uz (this model) | ASR transcript display normalization, mixed Uzbek/Russian |
Both models share the same ByT5 architecture. This variant is fine-tuned from the base with additional transcript-display objectives.
Quick Start
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
model_path = "islomov/rubai-corrector-transcript-uz"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForSeq2SeqLM.from_pretrained(model_path)
text = "bugun yaxshi kun. segodnya xoroshiy den."
inputs = tokenizer(f"correct: {text}", return_tensors="pt")
output_ids = model.generate(**inputs, max_new_tokens=256)
prediction = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
print(prediction)
Expected output:
Bugun yaxshi kun. Сегодня хороший день.
Real Example Outputs
The examples below are taken from this exact checkpoint's saved eval/test outputs.
Abbreviations / shorthand
Input: tlefon rqami
Output: Telefon raqami
Input: telefon rqami qaysi
Output: Telefon raqami qaysi
Apostrophes
Input: ozbekiston gozal mamlakat bolgan
Output: O'zbekiston go'zal mamlakat bo'lgan
Input: men ozim kordim
Output: Men o'zim ko'rdim.
OCR / ASR noise
Input: 0zbekiston Respub1ikasi
Output: O'zbekiston Respublikasi
Input: 5alom dostlar
Output: Salom do'stlar
Numbers
Input: uchrashuv o'n beshinchi yanvar kuni
Output: Uchrashuv 15-yanvar kuni
Input: narxi yigirma besh ming so'm
Output: Narxi 25 000 so'm
Mixed Uzbek + Russian
Input: bugun yaxshi kun. segodnya xoroshiy den.
Output: Bugun yaxshi kun. Сегодня хороший день.
Input: men bozorga bordim. tam ya kupil xleb.
Output: Men bozorga bordim. Там я купил хлеб.
Russian only
Input: segodnya xoroshaya pogoda
Output: Сегодня хорошая погода
Input: privet kak dela
Output: Привет как дела
Mixed script
Input: privet kak делa
Output: Привет как дела
Input: zaklad bersa keyin gaplashamiz
Output: Заклад bersa keyin gaplashamiz
Display-text cleanup
Input: mustahkamlik sinovida spark boshqa avtomobillarni ortda qoldirdi.
Output: Mustahkamlik sinovida Spark boshqa avtomobillarni ortda qoldirdi.
Input: kadrlarda kranning mashina old oynasi ustiga qulaganligini ko'rish mumkin
Output: Kadrlarda kranning mashina old oynasi ustiga qulaganligini ko'rish mumkin.
Known Tradeoff
This model is more display-oriented than the previous base.
That means:
- it is better at final punctuation and finished-sentence formatting
- it may add a final period in places where an old reference omitted it
Files
test_model.pysmall runnable example/test script for local use and HF packaging
How This Model Was Trained
This model is fine-tuned from rubai-corrector-base (which itself is built on google/byt5-small).
The fine-tuning added transcript-display objectives on top of the base correction capabilities:
- Uzbek transcript/display pairs from ASR output
- Russian recovery pairs from Latin-script ASR output
- Punctuation and formatting polish data
The model expects the correct: instruction prefix at inference time.
Acknowledgements
Special thanks to Davron Ibrokhimov for sponsoring this work and making it possible to keep these models open.
Thank you to the community that supports Uzbek language technology. In particular:
- MetaSell for support and resources
- Kotib for their support and collaboration on Uzbek STT
- Global Move for backing open Uzbek NLP work
Thanks to Arofat, Gulimshaxnoz, and many others who contributed in ways big and small. The list is too long to fit here, but every contribution matters and is appreciated.
Support my works and open-source movement: https://tirikchilik.uz/islomovs