0kam / sound_classifier

sound_classifier

Training and deploying deep sound classifiers, especially for bio-acoustic monitorings.

Installation

python3 -m pip install git+https://github.com/0kam/sound_classifier

Requirements

# For inference
tflite-runtime
or
tensorflow (tested on 2.10.0) # tflite-runtime has not support Python 3.10 yet, so use tensorflow.lite instead

# For a real-time inference using microphones
pyaudio
numpy
scipy # for audio resampling

# For training
tensorflow (tested on 2.10.0)
tensorflow-io
kapre
audiomentations
scikit-learn
tqdm
pandas
numpy
scypi
mlflow
omegaconf

Modules

• sound_classifier
  The main component of this package
  • core
    The core scripts for all models
    • audio_device.py
      For sampling audio from microphones
    • data.py
      For preparing audio datasets to train models
    • features.py
      For extracting features (such as Log-Mel spectrograms) from audios, using kapre package
    • sound_classifier.py
      An abstract based class for all classifier models
  • models
    Audio classifiers
    • yamnet.py
      A Google's YAMNet based audio classification model
    • fcn.py
      A small Fully Convolutional Neural network model
• zoo
  Implementation examples of models with sound_classifier
  Each directory contains Python scripts for training and inferencing the model. Also, some models supports TensorFlowLite inferencing for edge and mobile applications.
  • yamnet_google (!!under construction)
    The original YAMNet