Standardize Audio Training Inputs

Transform audio signals into 2D power spectra for net training inputs using a custom audio encoder

Open in Cloud
Download Notebook

Create an encoder that partitions the conformed audio data to transform one-dimensional audio data into a two-dimensional signal so that it can be used more effectively as input to a convolutional layer:

In[1]:=

enc=With[{innerEncoder=NetEncoder["Audio"]},NetEncoder[{"Function",Partition[Flatten@innerEncoder[#],512,512]&,{"Varying",512}}]]

Out[1]=

NetEncoder

Type:	Function
Output:	matrix (size: n ×512)



Apply the encoder to a recording of a bird:

In[2]:=

partionedData=enc[ExampleData[{"Audio","Bird"}]];

Plot the computed features:

In[3]:=

ListLinePlot[partionedData〚;;10〛]

Out[3]=

Create a different function encoder that standardizes the computed power spectrum of the training inputs according to a pre-determined distribution:

In[4]:=

enc=WithinnerEnc=NetEncoder["AudioSpectrogram"],mean=

,std=

,NetEncoder[{"Function",Map[(#1-mean)/std&,innerEnc[#]]&,{"Varying",201}}]

Out[4]=

NetEncoder

Type:	Function
Output:	matrix (size: n ×201)



Plot the features computed with the encoder:

In[5]:=

feats=enc[ExampleData[{"Audio","DogBark"}]];MatrixPlotTranspose@Normal@feats,

opts



Out[5]=

Related Symbols

NetEncoder
NetTrain
Audio

Publisher Information

Contributed by: Wolfram Staff

Wolfram Language Example Repository

Standardize Audio Training Inputs

See Also

Publisher Information