This Deep Learning model is for classifying audio using Tensorflow and Kapre (Keras Audio Preprocessors)
The development environment can be setup as follows:
conda create --name audio python==3.7
conda activate audio
git clone git@github.com:Patricknmaina/Deep-Audio-Classifier.git
cd Deep-Audio-Classifier/
pip install -r requirements.txt
To access the jupyter environment, run the following command:
jupyter-notebook
audio_clean.py is used to downsample the audio and remove dead cells in the audio with threshold detection using a signal envelope.
The file is run using the following command:
python scripts/audio_clean.py
In order to downsample the audio wavfiles by delta time, uncomment the split_wavs function. This will create a clean directory with the downsampled mono audio wavfiles.
In the train.py file, change the model_type to either conv1d, conv2d or lstm in the main function.
The delta time and sample rate should be constant from audio_clean.py
Then run the code as follows:
python scripts/train.py
The training process looks as follows:
After training the 3 models, the training logs are saved and implemented in notebooks/Model Plots.ipynb
The confusion matrix is developed in notebooks/Confusion Matrix and ROC.ipynb
The Receiver Operating Characteristic (ROC) curve is also developed in notebooks/Confusion Matrix and ROC.ipynb
In predict.py, change the model_type to either conv1d, conv2d or lstm in the main function to evaluate the performance of the trained models with test datasets.
Run the following code:
python scripts/predict.py
The prediction process is shown below:
This project was purely for experimental purposes. I will be experimenting with more updated wav files (possibly a different audio corpus), as well as more model architectures. So stay tuned!
Kapre -> for computation of audio transforms from time to frequency domain