Use kapre instead of this

kapre includes faster STFT/Melspectrogram with multi-channel supports as well as some other stuffs, works with both tensorflow and theano. For all cases kapre is better.

keras_STFT_layer

Do STFT & friends in Keras! For less hassle pre-processing

• [x] STFT
• [x] melgram

why

Because I am planning to compare the performance of some Convnets while changing parameters of STFT and storing all of them doesn't seem to make sense.

how

Theano backend only. image_dim_ordering() doesn't matter.

from stft import Spectrogram
import keras

len_src = 12000 * 8 # 8-second signal is your input
model = keras.Sequential()
specgram = Spectrogram(n_dft=512, n_hop=128, input_shape=(len_src, 1))
model.add(specgram)
model.add(BatchNormalization(axis=time_axis)) # recommended
model.add(your_awesome_network)
...

More info

• Jypyter notebook (STFT)
• Jypyter notebook (STFT)

Credits

I relied on Librosa codes.