We achieved 40th Place (Silver Medal). Congratulations to all of Team Members from BengalX: @iftekharamin , @mdfahimreshm , @fahimshahriarkhan π
\nI would like to acknowledge my Team Lead @iftekharamin Bhaiya for giving opportunity to do this competition and giving proper guideline for achieving silver medal.
\nDataset: We subset the dataset Based on @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.
\nData Cleaning: We filter audios which duration is less then 1 sec. as outlier and to not mislead model performance.
\nAugmentations: We used audio augmentations i.e, Noise , Background sound mixing, Speed up-down, SpecAug, Changing different Sampling Rates.
\nSTT Modeling: We used Indic wav2vec2 pretrained model. And we finetune with the Subset augmented dataset.
\nPost-processing- LM Decode: We used arijit indic pretrained KenLM.
\nPost-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}
\nPost-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
", "messages": [ { "id": "2497488", "postDate": "10/24/2023 16:43:34", "content": "We achieved 40th Place (Silver Medal). Congratulations to all of Team Members from BengalX: @iftekharamin , @mdfahimreshm , @fahimshahriarkhan π
\nI would like to acknowledge my Team Lead @iftekharamin Bhaiya for giving opportunity to do this competition and giving proper guideline for achieving silver medal.
\nDataset: We subset the dataset Based on @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.
\nData Cleaning: We filter audios which duration is less then 1 sec. as outlier and to not mislead model performance.
\nAugmentations: We used audio augmentations i.e, Noise , Background sound mixing, Speed up-down, SpecAug, Changing different Sampling Rates.
\nSTT Modeling: We used Indic wav2vec2 pretrained model. And we finetune with the Subset augmented dataset.
\nPost-processing- LM Decode: We used arijit indic pretrained KenLM.
\nPost-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}
\nPost-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
", "rawMarkdown": "We achieved 40th Place (Silver Medal). Congratulations to all of Team Members from **BengalX**: @iftekharamin , @mdfahimreshm , @fahimshahriarkhan π\n\n*I would like to acknowledge my Team Lead @iftekharamin Bhaiya for giving opportunity to do this competition and giving proper guideline for achieving silver medal.*\n\n**Dataset:** We subset the dataset Based on @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.\n\n**Data Cleaning:** We filter audios which duration is less then 1 sec. as outlier and to not mislead model performance.\n\n**Augmentations:** We used audio augmentations i.e, Noise , Background sound mixing, Speed up-down, SpecAug, Changing different Sampling Rates. \n\n**STT Modeling:** We used Indic wav2vec2 pretrained model. And we finetune with the Subset augmented dataset. \n\n**Post-processing- LM Decode:** We used arijit indic pretrained KenLM.\n\n**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} \n\n**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", "votes": null }, { "id": "2497717", "postDate": "10/24/2023 20:09:16", "content": "Great job! Congrats on the Silver Medal.
", "rawMarkdown": "Great job! Congrats on the Silver Medal.", "votes": null }, { "id": "2497842", "postDate": "10/25/2023 00:26:24", "content": "Thank you for sharing your solution.
\nWe couldn't make a spell checker work. How did it affectEd your LBand PB score?
Itβs improved 0.001 in LB & PB. Thank you for your comment and congratulations achieving 3rd place π
", "rawMarkdown": "Itβs improved 0.001 in LB & PB. Thank you for your comment and congratulations achieving 3rd place π", "votes": null } ], "comments": [ { "id": 2497717, "author_name": "scientiapotentia", "author_url": "", "post_date": "10/24/2023 20:09:16", "content": "Great job! Congrats on the Silver Medal.
", "votes": null, "replies": [] }, { "id": 2497842, "author_name": "ludditep", "author_url": "", "post_date": "10/25/2023 00:26:24", "content": "Thank you for sharing your solution.
\nWe couldn't make a spell checker work. How did it affectEd your LBand PB score?
Itβs improved 0.001 in LB & PB. Thank you for your comment and congratulations achieving 3rd place π
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2497488": "We achieved 40th Place (Silver Medal). Congratulations to all of Team Members from **BengalX**: @iftekharamin , @mdfahimreshm , @fahimshahriarkhan π\n\n*I would like to acknowledge my Team Lead @iftekharamin Bhaiya for giving opportunity to do this competition and giving proper guideline for achieving silver medal.*\n\n**Dataset:** We subset the dataset Based on @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.\n\n**Data Cleaning:** We filter audios which duration is less then 1 sec. as outlier and to not mislead model performance.\n\n**Augmentations:** We used audio augmentations i.e, Noise , Background sound mixing, Speed up-down, SpecAug, Changing different Sampling Rates. \n\n**STT Modeling:** We used Indic wav2vec2 pretrained model. And we finetune with the Subset augmented dataset. \n\n**Post-processing- LM Decode:** We used arijit indic pretrained KenLM.\n\n**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} \n\n**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", "2497717": "Great job! Congrats on the Silver Medal.", "2497842": "Thank you for sharing your solution.\nWe couldn't make a spell checker work. How did it affectEd your LBand PB score?", "2497895": "Itβs improved 0.001 in LB & PB. Thank you for your comment and congratulations achieving 3rd place π" }, "source": "meta" }