Elephant Sound Classification Using Raw Audio

Lightweight Machine Learning Models for Bioacoustic Monitoring

Associated Research Paper: 🔗 Elephant Sound Classification Using Deep Learning Optimization https://doi.org/10.3390/s25020352

Research Overview

This repository contains the full implementation, scripts, and datasets used in our study on raw-audio-based elephant vocalization classification, published in Sensors (MDPI), 2025.
The research focuses on developing efficient machine learning models that operate directly on raw waveforms, enabling deployment on resource-constrained edge devices for real-time wildlife monitoring.

We classify three elephant caller types:

• Rumble
• Roar
• Trumpet

Our results demonstrate that raw audio processing outperforms spectrogram-based pipelines, achieving higher accuracy while reducing computational overhead.

Dataset Availability

This project uses two sources of elephant vocalization data:

1. Classified Dataset (Included in This Repository)

A curated and pre-labeled dataset containing the three caller types:

• Rumble
• Roar
• Trumpet

Access the Dataset:
👉 Click here to open the dataset directory

The dataset is organized into:

dataset/
├── train/
├── val/
└── test/

All files are preprocessed and standardized to 6-second waveform clips.
This complete classified dataset is included directly within this GitHub repository.

Used Models

We experimented with the following models:

1. MobileNet V2:

   • Utilized pre-trained data and fine-tuned it on our dataset.

2. YAMNet:

   • Trained the model from scratch specifically for our datasets.

3. RawNet:

   • Trained the model from scratch for our datasets.

4. ElephantCallerNet based from ACDNet:

   • Trained the model from scratch for our datasets.

Pre-processing

The models are trained on raw audio files. However, the audio files need to be pre-processed before training. Here are the steps for pre-processing:

1. Dataset Division: The dataset is divided into test, train, and validation datasets, each containing the three classes.
2. Audio Duration: The models accept 6-second audio files.
   • If the waveform is greater than 6 seconds, it is trimmed to 6 seconds.
   • If the audio file is greater than 8 seconds, it is segmented into equal length sizes and padded or trimmed to make each segment 6 seconds.
   • Audio files with a duration less than 2 seconds are avoided.

Pipeline Diagram

               ┌────────────────────────┐
               │   Raw Audio Dataset    │
               │ (Rumble / Roar / Trumpet) 
               └──────────────┬─────────┘
                              │
                              ▼
                 ┌────────────────────────┐
                 │   Pre-processing       │
                 │  - Trim / Pad to 6 sec │
                 │  - Segment > 8 sec     │
                 │  - Remove < 2 sec      │
                 └──────────────┬─────────┘
                              │
                              ▼
          ┌────────────────────────────────────────┐
          │      Model Training (Raw Waveforms)     │
          │  MobileNet • YAMNet • RawNet • ElephantCallerNet
          └──────────────┬─────────────────────────┘
                              │
                              ▼
                 ┌────────────────────────┐
                 │       Evaluation        │
                 │  Accuracy & Metrics     │
                 └──────────────┬─────────┘
                              │
                              ▼
                 ┌────────────────────────┐
                 │       Inference         │
                 │ Predict caller type     │
                 └────────────────────────┘

Key Results

Raw Audio Classification Accuracy

Model	Training Approach	Input Type	Accuracy
MobileNet V2	Fine-tuned	Raw	82%
YAMNet	From scratch	Raw	78%
RawNet	From scratch	Raw	84%
ElephantCallerNet (Proposed)	From scratch	Raw	89%

---

Accuracy Comparison

A visual comparison of model accuracies across MobileNet, YAMNet, RawNet, and the proposed ElephantCallerNet, demonstrating the superior performance of raw-audio-based training.

---

Performance Comparison — Direct Raw Audio Classification

A detailed breakdown of performance metrics (accuracy) showing the advantages of processing raw waveforms over spectrogram-based approaches.

---

Elephant Monitoring System — Web Application

A system-level diagram of the end-to-end monitoring platform developed in this study, including audio capture, processing, classification, and dashboard visualization.

Environment Setup

To ensure a consistent environment for running the script, we have included an environment.yml file. This file can be used to set up a conda environment with all necessary dependencies.

Steps to Set Up the Environment

1. Install Conda:

   • If you don't have conda installed, download and install it from Anaconda or Miniconda.

2. Create the Environment:

   conda env create -f environment.yml

Usage Guide

The following steps describe how to process the raw audio files to ensure they meet the requirements of the machine learning models. The script segments, pads, or trims the audio files as necessary to produce 6-second audio files suitable for training.

Steps to Run the Audio Processing Script

1. Prepare the Script:

   • Save the provided script as process_audio.py.

2. Run the Script:

   • Run the script with the input and output directories as arguments:

   python process_audio.py <input_dir> <output_dir>

   Replace <input_dir> with the path to your directory containing the raw WAV files, and <output_dir> with the path where you want to save the processed audio files.

Example

Suppose you have a directory of raw audio files located at data/raw_audio and you want to save the processed audio files in data/processed_audio. You would run the script as follows:

python process_audio.py data/raw_audio data/processed_audio

This command processes each audio file in data/raw_audio, applies the necessary padding, segmentation, or trimming, and saves the processed files in data/processed_audio with appropriate naming.

Inference

After you have created or pre-processed your audio files, you can directly pass your audio file to inference code as follows;

python inference.py mobilenet /path/to/your/audio/file.wav

Replace 'mobilnet' with your preference model. This project is supporting 04 models, namely; Mobilnet, YAMNet, RawNet, and ElephantCallerNet

Citation

If you use this repository or the associated publication, please cite:

MDPI / ACS Style

Dewmini, H.; Meedeniya, D.; Perera, C. Elephant Sound Classification Using Deep Learning Optimization. Sensors 2025, 25, 352. https://doi.org/10.3390/s25020352

AMA Style

Dewmini H, Meedeniya D, Perera C. Elephant Sound Classification Using Deep Learning Optimization. Sensors. 2025; 25(2):352. https://doi.org/10.3390/s25020352

Chicago/Turabian Style

Dewmini, Hiruni, Dulani Meedeniya, and Charith Perera. 2025. "Elephant Sound Classification Using Deep Learning Optimization." Sensors 25, no. 2: 352. https://doi.org/10.3390/s25020352

APA Style

Dewmini, H., Meedeniya, D., & Perera, C. (2025). Elephant Sound Classification Using Deep Learning Optimization. Sensors, 25(2), 352. https://doi.org/10.3390/s25020352

BibTeX

@Article{s25020352,
  AUTHOR = {Dewmini, Hiruni and Meedeniya, Dulani and Perera, Charith},
  TITLE = {Elephant Sound Classification Using Deep Learning Optimization},
  JOURNAL = {Sensors},
  VOLUME = {25},
  YEAR = {2025},
  NUMBER = {2},
  ARTICLE-NUMBER = {352},
  URL = {https://doi.org/10.3390/s25020352},
  DOI = {10.3390/s25020352}
}