generate_noise_dataset.py

import numpy as np
import librosa
import pandas as pd
import numpy.random as rnd
import pathlib
import tqdm


dataset_folder = pathlib.Path('noise_data')
dataset_folder.mkdir(parents=True, exist_ok=True)

num_examples = 100

time_duration = 10
fs = 48000
num_samples = int(fs*time_duration)

prettyProgressBar = tqdm.tqdm(range(num_examples))

for i in prettyProgressBar:

    x = rnd.randn(num_samples)
    x /= np.maximum(np.max(x), -np.min(x))
    x *= rnd.rand()

    filepath = dataset_folder / f'noise_{i}.wav'
    librosa.output.write_wav(filepath, x, fs, norm=False)


file_list = list(dataset_folder.glob('*.wav'))
example_type = num_examples * ['eval']
scene_label = num_examples * ['not_scene']
city = num_examples * ['no_city']
identifier = num_examples * ['rand']
source_label = num_examples * ['z']

dict_data = {'filename': file_list, 
            'example_type': example_type,
            'scene_label': scene_label,
            'city': city,
            'identifier': identifier,
            'source_label': source_label}

df = pd.DataFrame(dict_data)

df.to_csv('noise_meta.csv', index=False, sep='\t')