SPELA - spectrogram layers
Rewrote kapre using tensorflow.keras
credits go to Keunwoo Choi for writing kapre
My main goal for rewriting it with tensorflow.keras is to use it with TensorFlow Lite
since Keunwoo Choi used core keras and I had problems converting the model to
tensorflow lite.
Implementing audio features inside the keras layers allows the preprocessing
computations to be done on the GPU as highlighted in their paper
Checkout this Speaker Recognition project to see the usage of Spela.
Installation
The package uses tensorflow but is not listed as requirement, please install it.
pip install spelaor
git clone https://github.com/kongkip/spela.git
cd spela
python setup.py installUsage
spectrogram
import tensorflow as tf
from spela.spectrogram import Spectrogram
# a one channel audio with 16000 sample rate
input_shape = (1, 16000)
x = get_data()
y = get_data()
model = tf.keras.Sequential()
model.add(Spectrogram(n_dft=512, n_hop=256, input_shape=(input_shape),
return_decibel_spectrogram=True, power_spectrogram=2.0,
trainable_kernel=False, name='static_stft'))
model.compile(optimizer=tf.keras.optimizers.Adam(lr=0.001), loss=tf.keras.losses.categorical_crossentropy
, metrics=["acc"])
print(model.summary())
model.fit(x,y)Mel Spectrogram
import tensorflow as tf
from spela.melspectrogram import Melspectrogram
# a one channel audio with 16000 sample rate
input_shape = (1, 16000)
x = get_data()
y = get_data()
model = tf.keras.Sequential()
model.add(Melspectrogram(sr=SR, n_mels=128,
n_dft=512, n_hop=256, input_shape=input_shape,
return_decibel_melgram=True,
trainable_kernel=False, name='melgram'))
model.compile(optimizer=tf.keras.optimizers.Adam(lr=0.001), loss=tf.keras.losses.categorical_crossentropy
, metrics=["acc"])
print(model.summary())
model.fit(x,y)
