SoundTrap Classispecies

Bioacoustic classifier written in Python

Features

Full classification pipeline from audio file to output class:

• Read audio files
• Analyse them, extracting a range of different features
• Train a classification model
• Evaluate the classification against test data
• Assess the classification output with different methods
• Output most likely class for each recording and relative probabilities
• Output matrix of all combination of parameters considered

Engineered to be extended:

• Plug-in your own feature extraction method
• Plug-in your own classification method
• Serialisation of intermediate steps, to avoid recomputing identical steps
• Object-oriented design
• Separation of model and controller
• Separation of settings configuration to customise all parameters in one place.

Engineered for performace:

• Multi-core operation that can process recordings, train and validate results in parallel (through the multiprocessing library)
• Dynamic comparison of different parameters (also saved to database and served in HTML)
• Logging of model operation with different levels of verbosity

Engineered for collaboration:

• Utilities for Comma-separated values (CSV) formatting of input data
• Utilities for conversion of extracted features to ARFF (Weka format).

Focus on results:

• Store results in a mongo database
• Serve results as a web page (HTML over configurable HTTP server)

Usage

This program classifies wildlife species according to a given model. A model is made of:

1. A feature extraction method
2. A classifier
3. A data set of sound files
4. A set of parameters [optional].

We provide a set of example models that can be copied for a quick start:

• orthoptera.py -- all dutch orthoptera
• birds.py -- a selection of 5 species of birds
• nfc-vs-dbc.py -- New Forest cicada and Roesel bush cricket.
• nips4b.py -- NIPS classification challenge
• ukorthoptera.py -- set of 28 orthoptera species in the UK.

orthoptera.py

all dutch orthoptera

birds.py

a selection of 5 species of birds

nfc-vs-dbc.py

New Forest cicada and Roesel bush cricket.

nips4b.py

NIPS classification challenge

ukorthoptera.py

set of 28 orthoptera species in the UK.

Existing features extraction methods are:

• mfcc -- Mel Frequency Cepstral Coefficients
• mel-spectrum -- Mel Spectrum
• hertz-spectrum -- Hertz Spectrum
• sp-kmeans -- Spherical k-means clustering

Existing supervised and unsupervised classifiers are:

• decisiontree -- Decision Tree Classifier
• randomforest -- Random Forest (Ensamble of Decision Trees)
• kmeans -- K-means clustering