The following project aims at implementing and training a neural network to perform speech denoising tasks given samples of speech signals and samples of noises.
This project was part of the MVA course "Apprentissage profond et traitement du signal, introduction et applications industrielles" given by Thomas Courtat.
The following points are covered:
- Generating a dataset by combining samples of speech and samples of noise.
- Debating the choice of the loss function for training and for evaluation.
- Reimplementing a ConvTasNet architecture.
- Benchmarking it against Asteroid’s pretrained model.