Similarity

Similarity between two songs have been defined by Tchernichovski et al (2000). This is a reimplementation of this algorithm.



In [1]:

    
%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import birdsonganalysis as bsa
from scipy.io import wavfile

Let's first reload two unrelated songs.



In [2]:

    
sr, sig = wavfile.read('../songs/simple.wav')
sr2, sig2 = wavfile.read('../songs/flashcam.wav')



In [3]:

    
spec_derivs_simple = bsa.spectral_derivs(sig)
fig = plt.figure(figsize=(16, 5))
ax = fig.gca()
ax = bsa.spectral_derivs_plot(spec_derivs_simple, 0.05, ax)
ax.set_title('Simple')
plt.show(fig)



In [4]:

    
spec_derivs_flash = bsa.spectral_derivs(sig2)
fig = plt.figure(figsize=(16, 5))
ax = fig.gca()
ax = bsa.spectral_derivs_plot(spec_derivs_flash, 0.01, ax)
ax.set_title('Flashcam')
plt.show(fig)

Compare their similarity

Let's see if simple could be similar to the reference song flashcam



In [5]:

    
sim1= bsa.similarity(sig, sig2, threshold=0.05)
print(sim1['similarity'])









    



0.349030574198



In [6]:

    
fig = bsa.similarity_plot(sim1, sig, sig2)
plt.show(fig)

As we can see, simple is similar to samba only by 36%.

Is simple similar to itself?



In [7]:

    
sim = bsa.similarity(sig, sig, threshold=0.05)
print(sim['similarity'])

1.0

Simple is similar to itself at 99%.



In [8]:

    
fig = bsa.similarity_plot(sim, sig, sig)
plt.show(fig)

Comparison between bells and a reproduction from a real bird!



In [2]:

    
sr, bells = wavfile.read('../songs/flashcam.wav')
sr, bells_rep = wavfile.read('../songs/flashcam_reproduction.wav')



In [3]:

    
spec_derivs_rep = bsa.spectral_derivs(bells_rep)
spec_derivs_bells = bsa.spectral_derivs(bells)



In [4]:

    
fig, (ax1, ax2) = plt.subplots(2, 1, figsize=(16, 10))
bsa.spectral_derivs_plot(spec_derivs_rep, 0.01, ax=ax1)
amp = bsa.song_amplitude(bells_rep)
amp[amp < 60] = np.nan
ax1 = bsa.plot_over_spec(amp, ax=ax1)
bsa.spectral_derivs_plot(spec_derivs_bells, 0.01, ax=ax2)
amp = bsa.song_amplitude(bells)
amp[amp < 60] = np.nan
ax2 = bsa.plot_over_spec(amp, ax=ax2)
plt.show(fig)



In [5]:

    
simrep = bsa.similarity(bells_rep, bells, threshold=0.05, silence_song_th=60, silence_ref_th=60)
print(simrep['similarity'])









    



0.538735294118



In [6]:

    
fig = bsa.similarity_plot(simrep, bells_rep, bells)
plt.show(fig)



In [ ]: