Keras Audio Library

Neural networks audio toolkit for Keras.

Running the library

Clone the repo using

git clone https://github.com/channelCS/keras_aud.git

and add the path to sys.

import sys
sys.path.insert(0, "path/to/keras_aud")

• Extracting audio features

from keras_aud import aud_audio
aud_audio.extract(feature_name, 'path/to/wavs', 'desired/features/path','yaml_file')

• Making a functional model

We make a functional CNN model='CNN' model in which we pass train_x having shape as (batch_size,1,dimx,dimy) and train_y with shape as (batch_size,num_classes).

from keras_aud import aud_model
miz=aud_model.Functional_Model(model='CNN',input_neurons=400,dropout1=0.1,
    act1='relu',act2='relu',act3=act3,nb_filter = 100, filter_length=5,
    num_classes=15,dimx=10,dimy=40)

This returns us a class object.

Parameters:

• dropout1: Dropout after every layer.
• act1: 1st Convolution layer activation.
• act2: 2nd Convolution layer activation.
• act3: Dense layer activation.
• nb_filter: Number of kernels.
• filter_length: Size of kernels.
• input_neurons: Number of neurons for the dense layer.

We can now fit this model and use the trained model for making predictions.

lrmodel=miz.prepare_model()
lrmodel.fit(train_x,train_y,batch_size=100,epochs=100,verbose=1)

• Extracting a single feature We use aud_feature to extract features a single wav file. We need to provide:

1. A path to the audio in .wav format
2. A yaml file containing feature description.
3. The feature we want to extract

We simply pass the required parameters to extract_one function.

from keras_aud import aud_feature
aud_feature.extract_one(feature_name = 'mel', wav_file = 'wavs/sample.wav', yaml_file = 'defaults.yaml')

• Plotting the feature

feature.plot_fig(X)

• Plotting the specgram

feature.plot_spec(X)