Research Software Jeantet · Çakır · Lontsi · Dufourq · MIT licence

Evolutionary Spectrogram Optimisation

A genetic algorithm for selecting informative mel-frequency bands in passive acoustic monitoring. The selected bands form a reduced-size input to a CNN classifier whose architecture is otherwise unchanged from the baseline.

Abstract

DomainPassive acoustic monitoring

TargetsHainan gibbon · Thyolo Alethe · Pin-tailed Whydah

HardwareLow-resource edge devices

Convolutional neural networks for bioacoustic classification typically operate on full mel-spectrograms produced after low-pass filtering and downsampling. The networks are large and the preprocessing pipeline is computationally expensive, limiting deployment on resource-constrained devices. This software implements ESO, an evolutionary algorithm that reduces the size of the network's input. A genetic algorithm searches over horizontal frequency bands of the unprocessed mel-spectrogram. Bands selected by the best chromosome are stacked or concatenated and used to train a CNN whose architecture is otherwise identical to the baseline. On the three datasets reported in the accompanying paper, ESO reduces mel-spectrogram size by 51 to 57 percent, trainable parameters by 64 to 72 percent, and energy consumption by 16 to 56 percent, while improving the F1-score by 1 to 6 percent.

1. Datasets

ESO was evaluated on three publicly available bioacoustic datasets that differ in target species, recording rate, soundscape complexity, and call structure.

Representative mel-spectrograms across the three study datasets. Panels show target vocalisations and the surrounding soundscape.

2. Representation

A gene encodes a single mel-spectrogram band by its lower frequency boundary \(P_k\) and height \(h_k\), with \(P_k \in [0, S_h - h_k]\), where \(S_h\) is the mel-spectrogram height. A chromosome is an ordered collection of genes and constitutes one candidate solution.

A chromosome with two genes applied to a mel-spectrogram of a Hainan gibbon vocalisation. Each gene defines a band by its position and height. Bands are extracted, then either stacked or concatenated.

3. Core abstractions

Five classes cover the algorithm. Each lives under eso/ga/ or eso/model/ and is documented in the API reference.

Gene

A single horizontal band, encoded as a position and a height. eso.ga.gene.Gene

Chromosome

An ordered set of genes and the CNN trained on the bands they define. eso.ga.chromosome.Chromosome

Population

A generation of chromosomes, evaluated and evolved together. eso.ga.population.Population

GeneticOperator

Reproduction, mutation, and crossover with user-defined rates. eso.ga.operator.GeneticOperator

SelectionOperator

Tournament selection. The tournament size is the only parameter. eso.ga.selection.SelectionOperator

Fitness

Relative F1 gain against the baseline plus relative reduction in trainable parameters. Chromosome.get_fitness

4. Documentation

Installation, configuration, and the algorithm's details, with cross-references to the source code.

• Installation

  System requirements, PyTorch wheel selection, and the editable install.

  Read →

• First run

  Settings JSON template and a minimal Python API example.

  Read →

• How ESO works

  Pipeline, representation, evolution, training, and evaluation.

  Read →

• Configuration

  Every field of the settings JSON, with types, defaults, and recommended values from the paper.

  Read →

• API reference

  Generated from docstrings in the eso package.

  Read →

• Citation

  BibTeX and venue.

  Read →