OpenSoundscape

OpenSoundscape is a utility library for analyzing bioacoustic data. It consists of Python modules for tasks such as preprocessing audio data, training machine learning models to classify vocalizations, estimating the spatial location of sounds, identifying which species' sounds are present in acoustic data, and more.

These utilities can be strung together to create data analysis pipelines. OpenSoundscape is designed to be run on any scale of computer: laptop, desktop, or computing cluster.

OpenSoundscape is currently in active development. If you find a bug, please submit an issue. If you have another question about OpenSoundscape, please email Sam Lapp (sam.lapp at pitt.edu).

Suggested Citation

Lapp, Rhinehart, Freeland-Haynes, Khilnani, Syunkova, and Kitzes, 2023. "OpenSoundscape v0.9.1".

Installation

OpenSoundscape can be installed on Windows, Mac, and Linux machines. It has been tested on Python 3.8, 3.9, 3.10, and 3.11. For Apple Silicon (M1 chip) users, Python >=3.9 is recommended and may be required to avoid dependency issues.

Most users should install OpenSoundscape via pip: pip install opensoundscape==0.9.1. Contributors and advanced users can also use Poetry to install OpenSoundscape.

For more detailed instructions on how to install OpenSoundscape and use it in Jupyter, see the documentation.

Features & Tutorials

OpenSoundscape includes functions to:

• load and manipulate audio files
• create and manipulate spectrograms
• train CNNs on spectrograms with PyTorch
• run pre-trained CNNs to detect vocalizations
• detect periodic vocalizations with RIBBIT
• load and manipulate Raven annotations
• estimate the location of sound sources from synchronized recordings

OpenSoundscape can also be used with our library of publicly available trained machine learning models for the detection of 500 common North American bird species.

For full API documentation and tutorials on how to use OpenSoundscape to work with audio and spectrograms, train machine learning models, apply trained machine learning models to acoustic data, and detect periodic vocalizations using RIBBIT, see the documentation.

Quick Start

Using Audio and Spectrogram classes

from opensoundscape import Audio, Spectrogram

#load an audio file and trim out a 5 second clip
my_audio = Audio.from_file("/path/to/audio.wav")
clip_5s = my_audio.trim(0,5)

#create a spectrogram and plot it
my_spec = Spectrogram.from_audio(clip_5s)
my_spec.plot()

Load audio starting at a real-world timestamp

from datetime import datetime; import pytz

start_time = pytz.timezone('UTC').localize(datetime(2020,4,4,10,25))
audio_length = 5 #seconds  
path = '/path/to/audiomoth_file.WAV' #an AudioMoth recording

Audio.from_file(path, start_timestamp=start_time,duration=audio_length)

Using a pre-trained CNN to make predictions on long audio files

from opensoundscape import load_model

#get list of audio files
files = glob('./dir/*.WAV')

#generate predictions with a model
model = load_model('/path/to/saved.model')
scores = model.predict(files)

#scores is a dataframe with MultiIndex: file, start_time, end_time
#containing inference scores for each class and each audio window

Training a CNN using audio files and Raven annotations

from sklearn.model_selection import train_test_split
from opensoundscape import BoxedAnnotations, CNN

# assume we have a list of raven annotation files and corresponding audio files
# load the annotations into OpenSoundscape
all_annotations = BoxedAnnotations.from_raven_files(raven_file_paths,audio_file_paths)

# pick classes to train the model on. These should occur in the annotated data
class_list = ['IBWO','BLJA']

# create labels for fixed-duration (2 second) clips 
labels = all_annotations.one_hot_clip_labels(
  cip_duration=2,
  clip_overlap=0,
  min_label_overlap=0.25,
  class_subset=class_list
)

# split the labels into training and validation sets
train_df, validation_df = train_test_split(labels, test_size=0.3)

# create a CNN and train on the labeled data
model = CNN(architecture='resnet18', sample_duration=2, classes=class_list)
model.train(train_df, validation_df, epochs=20, num_workers=8, batch_size=256)

Training a CNN with labeled audio data (one label per audio file):

from opensoundscape import CNN
from sklearn.model_selection import train_test_split

#load a DataFrame of one-hot audio clip labels
df = pd.read_csv('my_labels.csv') #index: paths; columns: classes
train_df, validation_df = train_test_split(df,test_size=0.2)

#create a CNN and train on 2-second spectrograms for 20 epochs
model = CNN('resnet18', classes=df.columns, sample_duration=2.0)
model.train(train_df, validation_df, epochs=20)
#the best model is automatically saved to a file `./best.model`