noise-guy is a program to generate audio music with added noise from various sources. By default, the audio sources used are members of the GTZAN dataset which sees popular use in the ML community for genre recognition. This program produces a set of noisy audio files and their original formats; this may be used to benchmark noise cancellation strategies.
Learn more about GTZAN here: https://www.kaggle.com/andradaolteanu/gtzan-dataset-music-genre-classification
Noise data that we use are taken from the BBC's sound effects library. Learn more about BBC sound effects and usage rights here: https://sound-effects.bbcrewind.co.uk/licensing
noise-guy will download our audio data and noise data from predetermined Amazon AWS S3 buckets. After this, the noise and audio are mixed using a python script. The output is a selection of noisy audio files from a range of music genres. This can be used for further processing.
To setup noise-guy you will need to download and unzip our datasets and to configure a python virtualenv to run our script. For convenience, a setup script is provided:
bash setup.sh
Use our running script to generate the data:
bash run.sh
Alternatively, noise-guy may be run manually:
source venv/bin/activate
python3 make-noisy-audio.py
After running, a selection of .wav files should be accessible in the ./output directory.
The make-noisy-audio.py script has a number of extra options for you to adjust if you would like more control or more data examples. These are accessible as command line arguments to the make-noisy-audio.py script:
--gtzan GTZAN directory of gtzan clean audio data for noise to be added
--noise NOISE directory of noise audio samples for adding to clean data
--out OUT output directory
--nr_clean NR_CLEAN number of clips to use from each gtzan genre
--nr_noise NR_NOISE number of noise clips to use
--rand RAND take a random audio file from each genre rather than the first