SunTwoShine/Zindi_Swahili_Audio_Classification

# Swahili Audio Classification (Zindi Challenge) Link to Zindi Challenge Page The goal of the Swahili Audio Classification challenge is to classify 12 swahili words recorded in wav-files. The words and english translations are shown below. | Swahili | English | | --- | --- | | ndio | yes | | hapana | no | | moja | one | | mbili | two | | tatu | three | | nne | four | | tano | five | | sita | six | | saba | seven | | nane | eight | | tisa | nine | | kumi | ten | A short introduction to the data is shown in the notebook "EDA.ipynb". 'Model_torchaudio.ipynb' contains the pytorch model including preprocessing, training and testing the model. A pretrained resnet18 model is used. This model can be either trained with normal spectrograms or mel spectrograms as preprocessing. Preprocessing using the mel spectrogram scored 0.168465426 (logloss) and reached **rank 16** of 40.