In this video, we learn one of many variations of Dynamic Time Warping (DTW) algorithm, namely ShapeDTW. Then, we will apply it to classify bird sounds. There will be 4000 recordings in the training dataset and 1000 recordings in the validation one. You can also see how I compute the ShapeDTW faster for all those recordings (2 hours vs 12+ hours).
00:00 Intro
01:23 Brief explanation of ShapeDTW
16:25 Bird sound data
17:27 Load libraries and the data
24:12 Audio feature extraction with MFCC
27:50 ShapeDTW between two sequences only
33:14 Very brief explanation of KNN
34:40 Computing ShapeDTW faster
39:04 Prediction on the validation dataset
40:49 Discussion
Link to the code: https://www.kaggle.com/code/leesstephanie/bird-song-classification-with-shapedtw/notebook
My 'final' version is version 9 (named 'final') and the one with 10 MFCC is on version 6 (named '10 MFCC, raw descriptor, symmetric1'. The one with only Derivative descriptor function is on version 3 (named 'Derivative shape descriptor').
Update: the final code is also available on my Github: https://github.com/stephanielees/BirdSoundClassification
#timeseries #classification #datascience #birdsounds #kaggle