SwishNet

PyTorch implementation of the SwishNet arquitecture.

arXiv:1812.00149: SwishNet: A Fast Convolutional Neural Network for Speech, Music and Noise Classification and Segmentation.

Usage

import torch

device = (
    "cuda"
    if torch.cuda.is_available()
    else "mps"
    if torch.backends.mps.is_available()
    else "cpu"
)

# MFCC frames as input: (batch_size, num_frames, num_coefficients)
x = torch.randn((1, 20, 32)).to(device)

model = SwishNet(in_channels=20, out_channels=2).to(device)
model(x)

In order to load a real audio file you can use torchaudio:

waveform, sample_rate = torchaudio.load("test.wav", normalize=True)
transform = transforms.MFCC(
    sample_rate=sample_rate,
    n_mfcc=20,
)

mfccs = transform(waveform).to(device)
model(mfccs)

alexcannan/swishnet-pytorch

SwishNet

Usage