Before into the topic, I would like to congratulate all team members: @nanaxing, @focuswilliam, @zhangjinru, @marcocheung0124. The first three of them are undergraduates and they are all new to Kaggle. CongratulationsBefore the topic, I would like to congratulate all team members: @nanaxing, @focuswilliam, @zhangjinru, @marcocheung0124. The first three of them are undergraduates and they are all new to Kaggle. Congratulation again on their first medal and the success of our first audio competition🥈🥈🥈!
\nThen I acknowledge @takanashihumbert for his published training notebook https://www.kaggle.com/code/takanashihumbert/bengali-sr-wav2vec-v1-bengali-training/notebook. Another acknowledgment is for @mbmmurad for his work to introduce the audios of the dataset https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0 in https://www.kaggle.com/code/mbmmurad/dataset-overlaps-with-commonvoice-11-bn/notebook.
\nDataset: We directly used the data provided by the organizer and did not use any external data.
\nTraining Environment: One Colab Pro+ account is utilized. Due to limitations of equipment and computing power, we are unable to train more models(like the punctuation model) and complete data sets.
\nData Augmentation: Some augmentations like HighLowPass, Noise, and PitchShift are used to increase robustness.
\nModel Training: Pretrained model is from https://www.kaggle.com/datasets/takanashihumbert/bengali-ex002. We put in one-tenth of the data for training each time (feature encoder and feature extractor take turns to freeze). For freezing the feature extractor, the lr is 2e-5 for warmup. For the feature encoder, it is 6e-6. In our experiments, bs=1. After training three-fifths of the data, the public score is 0.42 and private score is 0.503.
\nDecoding Parameters: We adjust the decoding parameters and displayed them in the notebook https://www.kaggle.com/code/yuliknormanowen/bengali-sr-wav2vec-v1-bengali-inference-for-v4?scriptVersionId=146374376.
\nThank you all so much for reading this. If you have any suggestions, we are happy to accept them. We would be very grateful if you could upvote this topic🥺🥺🥺🙏🙏
", "messages": [ { "id": "2498099", "postDate": "10/25/2023 06:24:12", "content": "Before into the topic, I would like to congratulate all team members: @nanaxing, @focuswilliam, @zhangjinru, @marcocheung0124. The first three of them are undergraduates and they are all new to Kaggle. CongratulationsBefore the topic, I would like to congratulate all team members: @nanaxing, @focuswilliam, @zhangjinru, @marcocheung0124. The first three of them are undergraduates and they are all new to Kaggle. Congratulation again on their first medal and the success of our first audio competition🥈🥈🥈!
\nThen I acknowledge @takanashihumbert for his published training notebook https://www.kaggle.com/code/takanashihumbert/bengali-sr-wav2vec-v1-bengali-training/notebook. Another acknowledgment is for @mbmmurad for his work to introduce the audios of the dataset https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0 in https://www.kaggle.com/code/mbmmurad/dataset-overlaps-with-commonvoice-11-bn/notebook.
\nDataset: We directly used the data provided by the organizer and did not use any external data.
\nTraining Environment: One Colab Pro+ account is utilized. Due to limitations of equipment and computing power, we are unable to train more models(like the punctuation model) and complete data sets.
\nData Augmentation: Some augmentations like HighLowPass, Noise, and PitchShift are used to increase robustness.
\nModel Training: Pretrained model is from https://www.kaggle.com/datasets/takanashihumbert/bengali-ex002. We put in one-tenth of the data for training each time (feature encoder and feature extractor take turns to freeze). For freezing the feature extractor, the lr is 2e-5 for warmup. For the feature encoder, it is 6e-6. In our experiments, bs=1. After training three-fifths of the data, the public score is 0.42 and private score is 0.503.
\nDecoding Parameters: We adjust the decoding parameters and displayed them in the notebook https://www.kaggle.com/code/yuliknormanowen/bengali-sr-wav2vec-v1-bengali-inference-for-v4?scriptVersionId=146374376.
\nThank you all so much for reading this. If you have any suggestions, we are happy to accept them. We would be very grateful if you could upvote this topic🥺🥺🥺🙏🙏
", "rawMarkdown": "Before into the topic, I would like to congratulate all team members: @nanaxing, @focuswilliam, @zhangjinru, @marcocheung0124. The first three of them are undergraduates and they are all new to Kaggle. CongratulationsBefore the topic, I would like to congratulate all team members: @nanaxing, @focuswilliam, @zhangjinru, @marcocheung0124. The first three of them are undergraduates and they are all new to Kaggle. Congratulation again on their first medal and the success of our first audio competition🥈🥈🥈!\n\nThen I acknowledge @takanashihumbert for his published training notebook [https://www.kaggle.com/code/takanashihumbert/bengali-sr-wav2vec-v1-bengali-training/notebook](url). Another acknowledgment is for @mbmmurad for his work to introduce the audios of the dataset [https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0](url) in [https://www.kaggle.com/code/mbmmurad/dataset-overlaps-with-commonvoice-11-bn/notebook](url).\n\n**Dataset**: We directly used the data provided by the organizer and did not use any external data.\n\n**Training Environment**: One Colab Pro+ account is utilized. Due to limitations of equipment and computing power, we are unable to train more models(like the punctuation model) and complete data sets. \n\n**Data Augmentation**: Some augmentations like HighLowPass, Noise, and PitchShift are used to increase robustness.\n\n**Model Training**: Pretrained model is from [https://www.kaggle.com/datasets/takanashihumbert/bengali-ex002](url). We put in one-tenth of the data for training each time (feature encoder and feature extractor take turns to freeze). For freezing the feature extractor, the lr is 2e-5 for warmup. For the feature encoder, it is 6e-6. In our experiments, bs=1. After training three-fifths of the data, the public score is 0.42 and private score is 0.503.\n\n**Decoding Parameters**: We adjust the decoding parameters and displayed them in the notebook [https://www.kaggle.com/code/yuliknormanowen/bengali-sr-wav2vec-v1-bengali-inference-for-v4?scriptVersionId=146374376](url). \n\nThank you all so much for reading this. If you have any suggestions, we are happy to accept them. We would be very grateful if you could upvote this topic🥺🥺🥺🙏🙏", "votes": null }, { "id": "2498176", "postDate": "10/25/2023 07:28:07", "content": "Congratulations!!!
", "rawMarkdown": "Congratulations!!!", "votes": null }, { "id": "2498263", "postDate": "10/25/2023 08:04:59", "content": "Congratulations! Really nice work! 🥰🥰
", "rawMarkdown": "Congratulations! Really nice work! 🥰🥰", "votes": null }, { "id": "2498655", "postDate": "10/25/2023 12:43:10", "content": "Congrats!✌ 🤗
", "rawMarkdown": "Congrats!✌ 🤗", "votes": null }, { "id": "2498783", "postDate": "10/25/2023 14:22:42", "content": "Congratulations! Really nice work @yuliknormanowen
", "rawMarkdown": "Congratulations! Really nice work @yuliknormanowen", "votes": null }, { "id": "2498865", "postDate": "10/25/2023 15:00:33", "content": "Congrats!✌ 🤗
", "rawMarkdown": "Congrats!✌ 🤗", "votes": null }, { "id": "2499337", "postDate": "10/26/2023 02:03:40", "content": "Thanks for sharing, congrats!
", "rawMarkdown": "Thanks for sharing, congrats!", "votes": null }, { "id": "2499684", "postDate": "10/26/2023 07:14:06", "content": "Nice working indeed in this competition @yuliknormanowen
", "rawMarkdown": "Nice working indeed in this competition @yuliknormanowen", "votes": null }, { "id": "2500064", "postDate": "10/26/2023 13:06:18", "content": "This is soo impressive, you must be really smart
", "rawMarkdown": "This is soo impressive, you must be really smart", "votes": null }, { "id": "2500274", "postDate": "10/26/2023 15:04:41", "content": "nice work!
", "rawMarkdown": "nice work!", "votes": null } ], "comments": [ { "id": 2498176, "author_name": "xiaowangiiiii", "author_url": "", "post_date": "10/25/2023 07:28:07", "content": "Congratulations!!!
", "votes": null, "replies": [] }, { "id": 2498263, "author_name": "rolianklay", "author_url": "", "post_date": "10/25/2023 08:04:59", "content": "Congratulations! Really nice work! 🥰🥰
", "votes": null, "replies": [] }, { "id": 2498655, "author_name": "nanaxing", "author_url": "", "post_date": "10/25/2023 12:43:10", "content": "Congrats!✌ 🤗
", "votes": null, "replies": [] }, { "id": 2498783, "author_name": "nithinreddy90", "author_url": "", "post_date": "10/25/2023 14:22:42", "content": "Congratulations! Really nice work @yuliknormanowen
", "votes": null, "replies": [] }, { "id": 2498865, "author_name": "focuswilliam", "author_url": "", "post_date": "10/25/2023 15:00:33", "content": "Congrats!✌ 🤗
", "votes": null, "replies": [] }, { "id": 2499337, "author_name": "artluo", "author_url": "", "post_date": "10/26/2023 02:03:40", "content": "Thanks for sharing, congrats!
", "votes": null, "replies": [] }, { "id": 2499684, "author_name": "tariqbashir", "author_url": "", "post_date": "10/26/2023 07:14:06", "content": "Nice working indeed in this competition @yuliknormanowen
", "votes": null, "replies": [] }, { "id": 2500064, "author_name": "maxmarriottclarke", "author_url": "", "post_date": "10/26/2023 13:06:18", "content": "This is soo impressive, you must be really smart
", "votes": null, "replies": [] }, { "id": 2500274, "author_name": "owenpauldeen", "author_url": "", "post_date": "10/26/2023 15:04:41", "content": "nice work!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2498099": "Before into the topic, I would like to congratulate all team members: @nanaxing, @focuswilliam, @zhangjinru, @marcocheung0124. The first three of them are undergraduates and they are all new to Kaggle. CongratulationsBefore the topic, I would like to congratulate all team members: @nanaxing, @focuswilliam, @zhangjinru, @marcocheung0124. The first three of them are undergraduates and they are all new to Kaggle. Congratulation again on their first medal and the success of our first audio competition🥈🥈🥈!\n\nThen I acknowledge @takanashihumbert for his published training notebook [https://www.kaggle.com/code/takanashihumbert/bengali-sr-wav2vec-v1-bengali-training/notebook](url). Another acknowledgment is for @mbmmurad for his work to introduce the audios of the dataset [https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0](url) in [https://www.kaggle.com/code/mbmmurad/dataset-overlaps-with-commonvoice-11-bn/notebook](url).\n\n**Dataset**: We directly used the data provided by the organizer and did not use any external data.\n\n**Training Environment**: One Colab Pro+ account is utilized. Due to limitations of equipment and computing power, we are unable to train more models(like the punctuation model) and complete data sets. \n\n**Data Augmentation**: Some augmentations like HighLowPass, Noise, and PitchShift are used to increase robustness.\n\n**Model Training**: Pretrained model is from [https://www.kaggle.com/datasets/takanashihumbert/bengali-ex002](url). We put in one-tenth of the data for training each time (feature encoder and feature extractor take turns to freeze). For freezing the feature extractor, the lr is 2e-5 for warmup. For the feature encoder, it is 6e-6. In our experiments, bs=1. After training three-fifths of the data, the public score is 0.42 and private score is 0.503.\n\n**Decoding Parameters**: We adjust the decoding parameters and displayed them in the notebook [https://www.kaggle.com/code/yuliknormanowen/bengali-sr-wav2vec-v1-bengali-inference-for-v4?scriptVersionId=146374376](url). \n\nThank you all so much for reading this. If you have any suggestions, we are happy to accept them. We would be very grateful if you could upvote this topic🥺🥺🥺🙏🙏", "2498176": "Congratulations!!!", "2498263": "Congratulations! Really nice work! 🥰🥰", "2498655": "Congrats!✌ 🤗", "2498783": "Congratulations! Really nice work @yuliknormanowen", "2498865": "Congrats!✌ 🤗", "2499337": "Thanks for sharing, congrats!", "2499684": "Nice working indeed in this competition @yuliknormanowen", "2500064": "This is soo impressive, you must be really smart", "2500274": "nice work!" }, "source": "meta" }