For research I chose this article "Phase-aware speech enchancement with Deep Complex U-Net", which describes the architecture and training of a convolutional neural network for improving speech, so-called denoising, and set up an experiment.
The main task is to develop a complex variation of the architecture of the well known UNet network to eliminate unwanted noise from the audio.
Its speciality and difference from other networks, such as SegNet, for semantic segmentation (not only that) lies in Skip-Connections and that the values of both input data and all of network parameters (convolution filters, etc.) are complex.
The main idea is that the early layers of the Encoder are concatenated with the " parallel " layers of the Decoder.
As a result of the convolution layers, we get a mask, which we multiply by the input time-frequency signal with noise and get a cleared time-frequency signal, which then passes the inverse Short-time Fourier transform.
For training we will use Noisy speech database for training speech enhancement algorithms and TTS models, which contains a data set for training and testing with 28 and 56 speakers in .wav audio files are 48 KHz. The 10-layer network architecture will be implemented, which looks like this:
A graph of changes in the value of the loss function during training and validation will be shown.
The PESQ metric will also be calculated.
Due to my lack of equipment with proper GPU (a laptop with 2 GB of GPU, so the model does not fit into the given memory, not to mention training) I had to consider alternatives for training:
- Training on Google Colab or another cloud service. Cloud services have strict session time limits, so it was decided to train on a small number of epochs.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent