Skip to content

README.md

Latest commit

 

History

History
83 lines (58 loc) · 2.32 KB

README.md

File metadata and controls

83 lines (58 loc) · 2.32 KB

Snore Source Separation for Denoising

Using UNet1D to learn how to remove non-snore noises from snoring audio signals

Overview

This repository contains all the necessary scripts to prepare, preprocess, and optionally train a UNet1D model for snore source separation or denoising. The approach involves:

  • Splitting and organizing raw data into Dataset/Raw (via prepareDataset.py).
  • Downsampling, normalizing, and augmenting the data for training (via preprocessDataset.py).
  • Training a UNet1D to remove noise from snoring signals (denoising).

Methods

Preprocessing

We used kaggle data as starting point and we applied minimal preprocessing involving:

  • normalization
  • downsampling to 16kHz

Architectures

We chose 2 well-known architectures for this tasks:

  • UNet1D [1]
  • CNNAutoEncoder [2]

No significant changes were applied to the original architectures.

Training

Later too lazy now

Inference

Later too

Results

Our strategy yields the following results

Model Similarity of Denoised Binary Classification
UNet1D [1] 76.88 100
CNNAutoEncoder [2] 67.50+- 100
WaveUNet1D [3] 83.17 100
ResUNet1D [4] xx.xx+-xx 100
AttentionUNet1D [5] xx.xx+-xx 100

How to run

Requirements

  • Download dataset from the Snoring Kaggle page
  • Place your dataset where you prefer
  • Create a virtual environment through python3 -m venv venv then source venv/bin/activate
  • Download dependencies by pip install -r requirements.txt
  • Optionally, create a wandb account and change the key wandb_entity on config.json file accordingly.

All the results were tested on a single NVIDIA RTX A5000 GPU.

Prepare and Preprocess data

python prepareDataset.py

python preprocessDataset.py

Train model for denoised model weight

python train_all.py

Inference weight for denoising

python inference_all.py

References

[1]

[2]