TechCabal Ewè Audio Translation Challenge

Introduction

This competition aims to evaluate an audio classification system for Ewe speakers in various contexts. The solution will be deployed on edge devices. For more details, please visit https://zindi.africa/competitions/techcabal-ewe-audio-translation-challenge.

My solution

I trained models using 3 following steps:

1. First, I trained a efficientnet-b0 model achieving 0.964 accuracy (public LB)
2. Second, I trained a mobilenet_v3_small model, reaching a 0.959 accuracy (public LB)
3. Third, since model size of the first model is bigger than 10 MB, but it has higher accuracy, so I decided to use pseudo-labels from test set and merge with training set to train the better mobilenet_v3_small model, which achieved 0.962 accuracy (public LB)

Feature engineering:

I limited audio time to 2 seconds and used Logmel features (n_mels = 128). With model 2 and 3, Delta features were used.

Augmentation:

MaskFreq

Folder structure

• dataset is folder of raw dataset, dataset/train includes 5335 wav files, dataset/test includes 2947 wav files.
• train_0.964.py is python script for training the first step.
• train_0.959.py is python script for training the second step.
• train_0.959_v2.py is python script for training the last step.
• saved_model is folder of models from all three steps
• result is folder of submissions. The result/Submission_cnn_0.962_from_0.959_v2.csv represents my final submission (0.962 public LB)
• inference.py: If you solely wish to execute inference, run this file!

Things I learnt from this competition

• I experimented various ways of feature engineering (MFCC, Logmels), but Logmels n_mels=128 gave me the best results. Decreasing n_mels resulted in decreased accuracy.
• I tried mobilenet_v4 but it not worked
• I also tried to develop my own CNN model, but the result is worsen.
• I explored other augmentation techniques like PitchShift, Add random noise but these did not improve results significantly.
• I used other ML training frameworks like Pytorch, huggingface, tf.keras, but none of them surpassed fastai

Thanks for this interesting challenge, This is my first time joining Zindi <3