• [Authors]
• [Introduction]
• [Dataset Description]
• [Solution Description]
• [Requirements]
• [Instalation]
• [Useful Links]

Deep Learning Experimentation for training an autoencoder model capable of removing noise features from given audios

We used the following datasets:

• MusicNet (https://www.kaggle.com/datasets/imsparsh/musicnet-dataset)

• Microsoft Scalable noisy speech Dataset (https://github.com/microsoft/MS-SNSD)

On the src folder you will find the notebooks used for building autoencoders:

• src\autoencoder1.ipynb : autoencoders trained on pure wav sequence
• src\autoencoder2.ipynb : autoencoders trained on mel spectrograms
• src\noise_adding.ipynb : notebook for overlaying noise audio on clean audio samples
• src\transformers.py : sample code for training transformers based on mel spectrograms

Basic reference of which libraries and versions were used

tensorboard=2.9.1
tensorboard-data-server=0.6.1
tensorboard-plugin-wit=1.8.1
tensorflow=2.11.0
tensorflow-estimator=2.9.0
tensorflow-intel=2.11.0
tensorflow-io-gcs-filesystem=0.30.0
librosa=0.10.0.post2
matplotlib-inline=0.1.6

• To create enviroment on conda:

conda create --name --file requirements.txt

• To create enviroment using pip

If you want a file which you can use to create a pip virtual environment (i.e. a requirements.txt in the right format) you can install pip within the conda environment, then use pip to create requirements.txt.

conda activate conda install pip pip freeze > requirements.txt

Then use the resulting requirements.txt to create a pip virtual environment:

python3 -m venv env source env/bin/activate pip install -r requirements.txt

https://www.tensorflow.org/io/tutorials/audio
https://towardsdatascience.com/audio-ai-isolating-instruments-from-stereo-music-using-convolutional-neural-networks-584ababf69de
https://www.kaggle.com/datasets/imsparsh/musicnet-dataset
https://www.tensorflow.org/tutorials/audio/simple_audio