This document summarizes work on classifying audio clips as containing sounds from North Atlantic right whales or not. Spectrograms of audio clips were generated and template matching and other feature engineering was used to extract over 450 features from the spectrograms. Random forests and gradient boosting models were trained on the features and achieved over 99% accuracy on test data using all features, with template matching features performing nearly as well. The goal is to help identify right whale calls to aid in collision avoidance for shipping lanes.
Report
Share
Report
Share
1 of 9
Download to read offline
More Related Content
Project Presentation (1)
1. March 11th, 2016
Jason Helgren, Jaime Pastor, Abhishek Singh
Have ye seen the Right Whale? A Classification Problem
2. Problem: Identifying the “Right Whale” Calls
Goal
Identify audio clips containing sounds from
the North Atlantic right whale.
Type
Binary classification
Data
30,000 2-second audio clips
Application
Collision avoidance in shipping lanes.
3. Audio Classification to Image Classification
Fourier Transformation
Frequency domain representation
of a time varying signal
Spectrogram
Frequency characteristics of a
signal over time
Audio Recording
Pressure variation in time. We
perceive these pressure variations as
sound
time
time
frequency
frequencydensitypressure
Idealized example - a tuning fork
5. Feature Engineering (I)
Processing images: Contrast-enhancement
Original spectrogram Processed image
Time and frequency metrics (300 features)
6. Feature Engineering (II)
Template matching (150 features)
Search for template over spectrogram and find the area of the image that
best matches the template
Original Image and matchWhale call
template example
Best match
(Correlation, x, y)
7. Modeling and results (I)
Model selection
❏ Random Forests
❏ Gradient Boosting
Hyper-parameter tuning
❏ Learning rate
❏ Sampling rate
❏ Number of trees
Scoring on the test set
❏ Using all variables (450 features)
❏ Using only template matching variables (150 features)
❏ Using non-template matching variables (300 features)
8. Modeling and results (II)
FEATURE SELECTION
TRAIN AUC
(24k obs.)
TEST AUC
(6k obs.)
All features
(450 features)
.9992 .9965
Only template matching features
(150 features)
.9982 .9951
Only non-template matching
features (300 features)
.9796 .9499
AUC Summary
9. Jason Helgren
jmhelgren@usfca.edu
in/jasonhelgren
“Have ye seen the Right Whale? A Classification Problem”
Kaggle 2013 Marinexplore Whale Detection Challenge
Project available at Github
Jaime Pastor
jpastorsamper@usfca.edu
in/jaimepastor2
Abhishek Singh
aasingh4@usfca.edu
in/abhishek19895
Thank you