About past torque prediction competition

 There was an interesting challenge to predict torque based on only audio recordings.

I've used 5-fold cv, score on local ~75, provisional 70.5, final ~72.
As there is small amount of data available - it's important to use augmentation and regularization.
As for augmentation I used cutout in time and frequency domain (audiomentations and specaugment libs).
As for the features I've splitted the raw-signal in 10-equal parts and calculated
statistical features for each: mean, std, kurtosis, skewness,
the same I've done for flattened mel-spectrogram,
also I've included category features.
The concatenated features vector was fed into xgboost and hypertuned. Also I tried to use nn on mel-spectrograms but got score ~68 on local only