A fully trainable Per-Channel Energy Normalisation layer for PyTorch. This repo will be in sporadic development, as there are a number of issues with the implementation that could be improved for usage with PyTorch. From a signal processing perspective however the layer is complete, the issues are primarily on the optimisation end.
See the following papers for details:
- Trainable frontend for robust and far-field keyword spotting
- Per-Channel Energy Normalization: Why and How
Navigate to repo and run:
pip install -e .`
from pcen import PCEN
# Fully Learnable PCEN layer
pcen = PCEN(
n_filters=40,
s_coef=0.05,
alpha=0.98,
delta=2.,
r_coef=2.,
trainable=True,
learn_s_coef=True,
per_channel_s=True
)
...
x = torchaudio.transforms.MelSpectrogram(n_mels=40)(audio)
x_pcen = pcen(x)`