# 44th Place Solution 🎉🎉🎉

Before into the topic, I would like to congratulate all team members:
[@nanaxing](https://www.kaggle.com/nanaxing)
,
[@focuswilliam](https://www.kaggle.com/focuswilliam)
,
[@zhangjinru](https://www.kaggle.com/zhangjinru)
,
[@marcocheung0124](https://www.kaggle.com/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](https://www.kaggle.com/nanaxing)
,
[@focuswilliam](https://www.kaggle.com/focuswilliam)
,
[@zhangjinru](https://www.kaggle.com/zhangjinru)
,
[@marcocheung0124](https://www.kaggle.com/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🥈🥈🥈!
Then I acknowledge
[@takanashihumbert](https://www.kaggle.com/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](https://www.kaggle.com/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)
.
Dataset
: We directly used the data provided by the organizer and did not use any external data.
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.
Data Augmentation
: Some augmentations like HighLowPass, Noise, and PitchShift are used to increase robustness.
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.
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)
.
Thank 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🥺🥺🥺🙏🙏