Speaker
Description
Multi-Resolution Neural Networks (MuReNN) are a new generation of models for deep learning for speech and audio processing. Compared to convolutional networks (convnets), they are more resource-efficient, less sensitive to initialization, and may generalize better across recording conditions. The key idea behind MuReNN is to learn a filterbank in which each filter is factorized between a non-learnable component (a complex-valued discrete wavelet) and a learnable component (a dilated convnet kernel). During this talk, we will present a differentiable and GPU-accelerated implementation of MuReNN in the PyTorch framework for Python. We will give a quick tutorial on how to build and train MuReNN layers and integrate them into full-fledged deep learning pipelines. Our open-source library is available at: https://github.com/kymatio/murenn