wxt808823/soundscape_ir

An open toolbox of soundscape information retrieval

soundscape_IR is a python-based toolbox of soundscape information retrieval, aiming to assist in the analysis of soundscape recordings. The toolbox is primarily desgined for: (1) visualization of soundscape dynamics (based on the MATLAB package Soundscape Viewer) and (2) audio source separation.

See https://meil-brcas-org.github.io/V1.1/index.html for technical documentation and more examples.

Installation

Dependencies:

• Python>=3.7
• numpy==1.21.6
• pandas==1.3.5
• audioread==2.1.9
• librosa==0.8.1
• scikit-learn==1.0.2
• scipy==1.7.3
• matplotlib==3.2.2
• plotly==5.5.0

To install soundscape_IR, clone the repository in your Python environment.

# Clone soundscape_IR from GitHub @meil-brcas-org
git clone https://github.com/meil-brcas-org/soundscape_IR.git

Then, install the requirements.txt in the package folder for installing required packages.

# Install required packages
cd soundscape_IR
pip install -r requirements.txt

Quick start

Audio visualization

soundscape_IR provides a function audio_visualization to transform an audio into a spectrogram on the hertz or mel scale. It also enables the use of Welch’s averaging method and spectrogram prewhitening in noise reduction. This example uses a short audio clip of sika deer calls and insect calls to demonstrate the ecoacoustic application of source separation.

from soundscape_IR.soundscape_viewer import audio_visualization

# Define spectrogram parameters
sound_train = audio_visualization(filename='case1_train.wav', path='./data/wav/', offset_read=0, duration_read=15,
                                  FFT_size=512, time_resolution=0.1, prewhiten_percent=10, f_range=[0,8000])

Model training

After preparing the training spectrogram, we can train a model with source_separation. NMF learns a set of basis functions to reconstruct the training spectrogram. In soundscape_IR, we can apply PC-NMF to separate the basis functions into two groups according to their source-specific periodicity. In this example, one group of basis funcitons is associated with deer call (mainly < 4 kHz) and another group is associated with noise (mainly > 3.5 kHz). Save the model for further applications.

from soundscape_IR.soundscape_viewer import source_separation

# Define model parameters
model=source_separation(feature_length=30, basis_num=10)

# Feature learning
model.learn_feature(input_data=sound_train.data, f=sound_train.f, method='PCNMF')

# Plot the basis functions of two sound source
model.plot_nmf(plot_type='W', source=1)
model.plot_nmf(plot_type='W', source=2)

# Save the model
model.save_model(filename='./data/model/deer_model.mat')

Deployment and spectrogram reconstruction

Generate another spectrogram for testing the source separation model.

# Prepare a spectrogram
sound_predict=audio_visualization(filename='case1_predict.wav', path='./data/wav/', offset_read=30, duration_read=15,
                                    FFT_size=512, time_resolution=0.1, prewhiten_percent=10, f_range=[0,8000])

Load the saved model and perform source separation. After the prediction procedure, plot the reconstructed spectrograms to evaluate the separation of deer calls and noise.

# Deploy the model
model=source_separation()
model.load_model(filename='./data/model/deer_model.mat')
model.prediction(input_data=sound_predict.data, f=sound_predict.f)

# View individual reconstructed spectrogram
model.plot_nmf(plot_type = 'separation', source = 1)
model.plot_nmf(plot_type = 'separation', source = 2)

Presence detection

With the reconstructed spectrogram, we can use the function spectrogram_detection to detect the presence of target signals (e.g., deer calls). This function will generate a txt file contains the beginning time, ending time, minimum frequency, and maximum frequency of each detected call. Explore the detection result in Raven software.

from soundscape_IR.soundscape_viewer import spectrogram_detection

# Choose the source for signal detection
source_num=2
    
# Define the detection parameters
sp=spectrogram_detection(model.separation[source_num-1], model.f, threshold=5.5, smooth=1, minimum_interval=0.5, 
                           filename='deer_detection.txt', path='./data/txt/')

More tutorials

• Demo of audio source separation - Detecting deer calls from tropical forest soundscapes
• Demo of audio source separation - Learning the diversity of underwater sounds from subtropical estuary soundscapes
• Demo of batch processing - Automatically analyzing a large number of soundscape recordings
• Demo of Soundscape Viewer - Investigating soundscape dynamics via visualization of long-duration recordings, blind source separation, and clustering

Currently ongoing developments

• Soundscape spatial analysis
• Plotly-based interactive plots

Future works

• GPU accelaration

Citing this work

If you find this package useful in your research, we would appreciate citations to:

• Sun, Y-J, Yen, S-C, & Lin, T-H (2022). soundscape_IR: A source separation toolbox for exploring acoustic diversity in soundscapes. Methods in Ecology and Evolution, 13: 2347-2355. https://doi.org/10.1111/2041-210X.13960

Bugs report and suggestions

If you encounter any bug or issue, please contact Dr. Tzu-Hao Lin via [email protected]. Suggestions are also appreciated!

About the team

Marine Ecoacoustics and Informatics Lab (MEIL)
Led by Dr. Tzu-Hao Lin, the MEIL investigates the applications of ecological informatics in biodiversity monitoring and conservation management. If you're interested in our work, please check our website or follow us on facebook.