Top 4% solution for Freesound General-purpose Audio tagging Challenge

This was my first time to handle sound classification problem because I missed a similar competition from kaggle. It is sad that I didn't understand this is a research competition which means there was no medals althrough my solution ranked 19th. I think that was why not so many people joint it. But after all, I learnt to deal similar problems.

Bad begin

It was natual to use RNN architectures since sound is a time serials signal, plus the data provided has different length which can only be handled in a whole by RNNs. But I got no good results after trying LSTMs and GRUs no matter how deep they were.

Then I tried to trim the sounds to the first 5 seconds and input them into a 1D CNNs. Then I even tried to transform the sound into mfcc 2d images and trained a 2d classifier on them. Still, none of above methods got really god result.

After search on internet and the kaggle forum, I got no good new ideas.

Master's Insight

The blog from kaggle's global 1st bestfitting who got the 1st place in only 1 year gave me a good suggest: Read up all best solutions from the previous competitions. This is a really gold suggestion which turns kaggle from a competitions platform into the best machine learning library in the world. I found a very similar competition just a few month ago named TensorFlow Speech Recognition Challenge in which the winner turned the sounds into mel spectogram 2d images and then trained a image classifier on them.

I followed this method and got a better score in leaderboard. I got even better scores after I started training from pretrained weights from keras. From this competitions, I really began to realize that the quality of the starting pretrained weights matters a lot in all kinds of competitions.

Stack Them All

The best solution from me is to train a big network contained as many models from many formats of sound including:

1. Image classifier on mel spectograms
2. Image classifier mfcc spectograms;
3. 1D CNNs on first 5 seconds;
4. GBTs on statistical features from sounds;

I really had fun in doing this since it is my first time to implement stack and it really worked.