def inject_noise(data, noise_color, noise_level):
noise_data = noise(len(data), color=noise_color)
noise_energy = np.sqrt(noise_data.dot(noise_data) / noise_data.size)
data_energy = np.sqrt(data.dot(data) / data.size)
data += noise_level * noise_data * data_energy / noise_energy
return data
"""
(Monofractal) Detrended Fluctuation Analysis example
"""
import numpy as np
from acoustics.generator import noise
import matplotlib.pyplot as plt
n = 8000
whitenoise = noise(n, color='white')
pinknoise = noise(n, color='pink')
# fractional brownian noise
RW1 = np.cumsum(whitenoise - np.mean(whitenoise))
RW2 = np.cumsum(pinknoise - np.mean(pinknoise))
# root-mean-square variation of time series (typical TS approach)
## average variation in time series
## average amplitude is similar for two very different signals
RMS_white = np.sqrt(np.mean(np.power(whitenoise, 2)))
RMS_pink = np.sqrt(np.mean(np.power(pinknoise, 2)))
print("Average variation for white noise {} and pink noise {}".format(
RMS_white, RMS_pink))
# Local Detrending: local RMS variation in the time series
ts = pinknoise
X = np.cumsum(ts - np.mean(ts))
scale = 1000
m = 1
segments = int(np.floor(len(X) / scale))
index = []
fit = []
def get_noise(length: int) -> np.ndarray:
color = random.choice(['white', 'pink', 'blue', 'brown', 'violet'])
return noise(length, color).astype('float32')
from acoustics.generator import noise
from scipy.io import wavfile as wf
from os.path import join as jp
import numpy as np
if __name__ == '__main__':
noise_dir = jp('data', 'train', 'audio', '_background_noise_')
white_noise_fn = jp(noise_dir, 'custom_white_noise.wav')
# 'white' and 'pink' already exist
noise_colors = ['blue', 'brown', 'violet']
for noise_color in noise_colors:
noise_fn = jp(noise_dir, 'custom_%s_noise.wav' % noise_color)
wf.write(
noise_fn, 16000,
np.int16(((noise(16000 * 60, color=noise_color)) / 3) * 32767))
print("Done!")
def noise_generator(kind, duration, frame_rate=44100, amplitude=0.5):
from acoustics.generator import noise
samples = noise(frame_rate, kind)
samples = [s / 5 * amplitude for s in samples]
return cycle_n(samples, int(duration * frame_rate / 1000))
step_size = 1
q = 2
order = 1
XY = []
FBM = []
FWN = []
#colors = ['white', 'white','pink','brown']
colors = ['white', 'white', 'pink']
process_label = [
'$Anti$-$persistent$ $process$', '$Short$-$range$ $correlations$',
'$Persistent$ $process$'
]
for i, c in enumerate(colors):
fwn = noise(n, color=c)
# create some colored fBm data
if i == 0:
fbm = np.mat(fwn) + (np.mat(np.cumsum(fwn) - np.mean(fwn))) / 5
else:
fbm = np.mat(np.cumsum(fwn) - np.mean(fwn))
uneven = fbm.shape[1] % 2
if uneven:
fbm = fbm[0, :-1]
# AFA step
XY.append(md.multi_detrending(fbm, step_size, q, order))
FBM.append(fbm)
FWN.append(fwn)
FWN[0] = difference(FWN[0], interval=1)
"""
write time series to csv for cross-platform testing
"""
import csv
import numpy as np
from acoustics.generator import noise
n = 8000
ts = noise(n, color = 'white')
print(ts)
with open('ts.csv', 'w') as f:
wr = csv.writer(f)
wr.writerow(ts)
def test_power_density(self, color, samples, power_density_change):
fs = 48000
_, L = octaves(noise(samples, color), fs, density=True)
change = np.diff(L).mean()
assert (np.abs(change - power_density_change) < self.ERROR)
def test_length(self, color, samples):
assert (len(noise(samples, color)) == samples)
def apply_noise(audio, color, snr):
noise_data = noise(len(audio.samples), color=color)
return _mix(audio, noise_data, snr)
def test_power_density(self, color, samples, power_density_change):
fs = 48000
_, L = octaves(noise(samples, color), fs, density=True)
change = np.diff(L).mean()
assert (np.abs(change - power_density_change) < self.ERROR)
def test_length(self, color, samples):
assert (len(noise(samples, color)) == samples)
