In [3]:
%matplotlib inline
import seaborn
import numpy, scipy, matplotlib.pyplot as plt, IPython.display as ipd
import librosa, librosa.display
plt.rcParams['figure.figsize'] = (14, 5)
Harmonic-Percussive Source Separation¶
Load two files: one harmonic, and one percussive.
In [24]:
xh, sr_h = librosa.load('audio/prelude_cmaj.wav', duration=7, sr=None)
In [25]:
ipd.Audio(xh, rate=sr)
Out[25]:
In [26]:
xp, sr_p = librosa.load('audio/125_bounce.wav', duration=7, sr=None)
In [27]:
ipd.Audio(xp, rate=sr)
Out[27]:
In [28]:
print len(xh), len(xp)
In [29]:
print sr_h, sr_p
Add the two signals together, and rescale:
In [30]:
x = xh/xh.max() + xp/xp.max()
x = 0.5 * x/x.max()
In [32]:
x.max()
Out[32]:
Listen to the combined audio signal:
In [34]:
ipd.Audio(x, rate=sr_h)
Out[34]:
Compute the STFT:
In [35]:
X = librosa.stft(x)
Take the log-ampllitude for display purposes:
In [36]:
Xmag = librosa.amplitude_to_db(X)
Display the log-magnitude spectrogram:
In [38]:
librosa.display.specshow(Xmag, sr=sr_h, x_axis='time', y_axis='log')
plt.colorbar()
Out[38]:
Perform harmonic-percussive source separation:
In [39]:
H, P = librosa.decompose.hpss(X)
Compute the log-amplitudes of the outputs:
In [41]:
Hmag = librosa.amplitude_to_db(H)
Pmag = librosa.amplitude_to_db(P)
Display each output:
In [42]:
librosa.display.specshow(Hmag, sr=sr_h, x_axis='time', y_axis='log')
plt.colorbar()
Out[42]:
In [43]:
librosa.display.specshow(Pmag, sr=sr_p, x_axis='time', y_axis='log')
plt.colorbar()
Out[43]:
Transform the harmonic output back to the time domain:
In [44]:
h = librosa.istft(H)
Listen to the harmonic output:
In [45]:
ipd.Audio(h, rate=sr_h)
Out[45]:
Transform the percussive output back to the time domain:
In [46]:
p = librosa.istft(P)
Listen to the percussive output:
In [47]:
ipd.Audio(p, rate=sr_p)
Out[47]: