# 40th Place Solution without External Dataset!

We achieved 40th Place (Silver Medal). Congratulations to all of Team Members from
BengalX
:
[@iftekharamin](https://www.kaggle.com/iftekharamin)
,
[@mdfahimreshm](https://www.kaggle.com/mdfahimreshm)
,
[@fahimshahriarkhan](https://www.kaggle.com/fahimshahriarkhan)
🎉
I would like to acknowledge my Team Lead
[@iftekharamin](https://www.kaggle.com/iftekharamin)
Bhaiya for giving opportunity to do this competition and giving proper guideline for achieving silver medal.
Dataset:
We subset the dataset Based on
[@imtiazprio](https://www.kaggle.com/imtiazprio)
published train metadata. We used condition to filter clean dataset from train meta features (yellowking_preds & google_preds) wer = 90% similar. After this we further filter dataset which mos_pred > 2. And we found around 100k+ datapoints.
Data Cleaning:
We filter audios which duration is less then 1 sec. as outlier and to not mislead model performance.
Augmentations:
We used audio augmentations i.e, Noise , Background sound mixing, Speed up-down, SpecAug, Changing different Sampling Rates.
STT Modeling:
We used Indic wav2vec2 pretrained model. And we finetune with the Subset augmented dataset.
Post-processing- LM Decode:
We used arijit indic pretrained KenLM.
Post-processing-Punctuation:
We used xashru/punctuation-restoration repo with xlm-roberta-base model and fine tune this competition dataset as punctuation restoration. We only consider 4 punctuation classes : {'O': 0, 'COMMA': 1, 'PERIOD': 2, 'QUESTION': 3}
Post-processing-Erro Correction:
We used this repo solution as further error correction
[https://github.com/Tawkat/Bengali-Spell-Checker-and-Auto-Correction-Suggestion-for-MS-Word](https://github.com/Tawkat/Bengali-Spell-Checker-and-Auto-Correction-Suggestion-for-MS-Word)