Remove background noise with signal processing tools

Environmental audio recordings usually have stationary noise that needs to be removed to enhance the signal to noise ratio of biological sounds. This example shows different ways to remove stationary background noise using spectral subtraction techniques. These techniques are applied over the spectrogram and return a 2D matrix. We use the sharpness metric to have a quantitative estimation of how well is the noise reduction. For a more comprehensive analysis, other metrics should be use in complement.

# sphinx_gallery_thumbnail_path = './_images/sphx_glr_plot_remove_background_002.png'

Load required modules

from maad.util import plot2d, power2dB, dB2power
from maad.sound import (load, spectrogram,
                       remove_background, median_equalizer,
                       remove_background_morpho,
                       remove_background_along_axis,
                       pcen)
from maad.sound import spectral_snr
import numpy as np

from timeit import default_timer as timer

import matplotlib.pyplot as plt

Load and plot the spectrogram of the original audio file

First, we load the audio file and take its spectrogram. The linear spectrogram is then transformed into dB. The dB range is 96dB which is the maximum dB range value for a 16bits audio recording. We add 96dB in order to get have only positive values in the spectrogram.

s, fs = load('../../data/tropical_forest_morning.wav')
Sxx, tn, fn, ext = spectrogram(s, fs, fcrop=[0,20000], tcrop=[0,60])
Sxx_dB = power2dB(Sxx, db_range=96) + 96
print("SNR = %2.3f dB" % spectral_snr(Sxx)[2])

plot2d(Sxx_dB, extent=ext, title='original',
       vmin=np.median(Sxx_dB), vmax=np.median(Sxx_dB)+40)

SNR = 3.505 dB

Test different methods to remove stationary background noise

Test the function “remove_background”

start = timer()
X1, noise_profile1, _ = remove_background(Sxx_dB)
elapsed_time = timer() - start
print("---- test remove_background -----")
print("duration %2.3f s" % elapsed_time)
print("SNR = %2.3f dB" % spectral_snr(dB2power(X1))[2])

plot2d(X1, extent=ext, title='remove_background',
       vmin=np.median(X1), vmax=np.median(X1)+40)

---- test remove_background -----
duration 0.093 s
SNR = 3.084 dB

Test the function “median_equalizer”

start = timer()
X2 = median_equalizer(Sxx)
X2 = power2dB(X2)
elapsed_time = timer() - start
print("---- test median_equalizer -----")
print("duration %2.3f s" % elapsed_time)
print("SNR = %2.3f dB" % spectral_snr(dB2power(X2))[2])

plot2d(X2,extent=ext, title='median_equalizer',
       vmin=np.median(X2), vmax=np.median(X2)+40)

---- test median_equalizer -----
duration 0.270 s
SNR = 7.032 dB

Test the function “remove_background_morpho”

start = timer()
X3, noise_profile3,_ = remove_background_morpho(Sxx_dB, q=0.95)
elapsed_time = timer() - start
print("---- test remove_background_morpho -----")
print("duration %2.3f s" % elapsed_time)
print("SNR = %2.3f dB" % spectral_snr(dB2power(X3))[2])

plot2d(X3, extent=ext, title='remove_background_morpho',
       vmin=np.median(X3), vmax=np.median(X3)+40)

---- test remove_background_morpho -----
duration 1.334 s
SNR = 2.074 dB

Test the function “remove_background_along_axis”

start = timer()
X4, noise_profile4 = remove_background_along_axis(Sxx_dB,mode='median', axis=1)
elapsed_time = timer() - start
print("---- test remove_background_along_axis -----")
print("duration %2.3f s" % elapsed_time)
print("SNR = %2.3f dB" % spectral_snr(dB2power(X4))[2])

plot2d(X4,  extent=ext, title='remove_background_along_axis',
       vmin=np.median(X4), vmax=np.median(X4)+40)

plt.tight_layout()

---- test remove_background_along_axis -----
duration 0.115 s
SNR = 5.758 dB

Test the function “pcen”

start = timer()
X5, noise_profile5, PCENxx = pcen(Sxx, gain=0.1, bias=5, power=0.25, b=0.01, eps=1e-4)
X5 = power2dB(X5, db_range=96) + 96
elapsed_time = timer() - start
print("---- test pcen -----")
print("duration %2.3f s" % elapsed_time)
print("SNR = %2.3f dB" % spectral_snr(dB2power(X5))[2])

plot2d(X5,  extent=ext, title='Per Channel Energy Normalization (PCEN)',
       vmin=np.median(X5), vmax=np.median(X5)+40)

plt.tight_layout()

---- test pcen -----
duration 0.180 s
SNR = 3.473 dB