sound_annalysis/cover_song_identification.py

## tutorial from: https://essentia.upf.edu/essentia_python_examples.html
#
##Cover song identification (CSI) in MIR is a task of identifying when two musical recordings are derived from the same music composition. The cover of a song can be drastically different from the original recording. It can change key, tempo, instrumentation, musical structure or order, etc.
##
##Essentia provides open-source implmentation of some state-of-the-art cover song identification algorithms. The following process-chain is required to use this CSI algorithms.
##
##    Tonal feature extraction. Mostly used by chroma features. Here we use HPCP.
##
##    Post-processing of the features to achieve invariance (eg. key) [3].
##
##    Cross similarity matrix computation ([1] or [2]).
##
##    Local sub-sequence alignment to compute the pairwise cover song similarity distance [1].
##
##In this tutorial, we use HPCP, ChromaCrossSimilarity and CoverSongSimilarity algorithms from essentia.

import essentia.standard as estd
from essentia.pytools.spectral import hpcpgram

import IPython

yesterday_original = 'audio/Yesterday (Remastered 2009).mp3'
yesterday_cover_01 = 'audio/Yesterday - The Beatles - Connie Talbot (Cover).mp3'
different_song = 'audio/Jacques Brel - Ne Me Quitte Pas.mp3'

IPython.display.Audio(yesterday_original)
IPython.display.Audio(yesterday_cover_01)
IPython.display.Audio(different_song)

# query cover song
original_song = estd.MonoLoader(filename=yesterday_original, sampleRate=32000)()

true_cover_01 = estd.MonoLoader(filename=yesterday_cover_01, sampleRate=32000)()

# wrong match
false_cover_1 = estd.MonoLoader(filename=different_song, sampleRate=32000)()

## Now let’s compute Harmonic Pitch Class Profile (HPCP) chroma features of these audio signals.

query_hpcp = hpcpgram(original_song, sampleRate=32000)

true_cover_hpcp_1 = hpcpgram(true_cover_01, sampleRate=32000)

false_cover_hpcp = hpcpgram(false_cover_1, sampleRate=32000)

## plotting the hpcp features
#%matplotlib inline
import matplotlib.pyplot as plt

fig = plt.gcf()
fig.set_size_inches(14.5, 4.5)

plt.title("Query song HPCP")
plt.imshow(query_hpcp[:500].T, aspect='auto', origin='lower', interpolation='none')

## Next steps are done using the essentia ChromaCrossSimilarity function,
## 
##     Stacking input features
## 
##     Key invariance using Optimal Transposition Index (OTI) [3].
## 
##     Compute binary chroma cross similarity using cross recurrent plot as described in [1] or using OTI-based chroma binary method as detailed in [3]

crp = estd.ChromaCrossSimilarity(frameStackSize=9,
                                 frameStackStride=1,
                                 binarizePercentile=0.095,
                                 oti=True)

true_pair_crp_1 = crp(query_hpcp, true_cover_hpcp_1)

fig = plt.gcf()
fig.set_size_inches(15.5, 5.5)

plt.title('Cross recurrent plot [1]')
plt.xlabel('Yesterday accapella cover')
plt.ylabel('Yesterday - The Beatles')
plt.imshow(true_pair_crp_1, origin='lower')





## Compute binary chroma cross similarity using cross recurrent plot of the non-cover pairs

crp = estd.ChromaCrossSimilarity(frameStackSize=9,
                                 frameStackStride=1,
                                 binarizePercentile=0.095,
                                 oti=True)

false_pair_crp = crp(query_hpcp, false_cover_hpcp)

fig = plt.gcf()
fig.set_size_inches(15.5, 5.5)

plt.title('Cross recurrent plot [1]')
plt.xlabel('Come together cover - Aerosmith')
plt.ylabel('Yesterday - The Beatles')
plt.imshow(false_pair_crp, origin='lower')


## Alternatively, you can also use the OTI-based binary similarity method as explained in [2] to compute the cross similarity of two given chroma features.

csm = estd.ChromaCrossSimilarity(frameStackSize=9,
                                 frameStackStride=1,
                                 binarizePercentile=0.095,
                                 oti=True,
                                 otiBinary=True)

oti_csm = csm(query_hpcp, false_cover_hpcp)

fig = plt.gcf()
fig.set_size_inches(15.5, 5.5)

plt.title('Cross similarity matrix using OTI binary method [2]')
plt.xlabel('Come together cover - Aerosmith')
plt.ylabel('Yesterday - The Beatles')
plt.imshow(oti_csm, origin='lower')




##     Finally, we compute an asymmetric cover song similarity measure from the pre-computed binary cross simialrity matrix of cover/non-cover pairs using various contraints of smith-waterman sequence alignment algorithm (eg. serra09 or chen17).
## 
## Computing cover song similarity distance between ‘Yesterday - accapella cover’ and ‘Yesterday - The Beatles’

score_matrix, distance = estd.CoverSongSimilarity(disOnset=0.5,
                                                  disExtension=0.5,
                                                  alignmentType='serra09',
                                                  distanceType='asymmetric')(true_pair_crp_1)

fig = plt.gcf()
fig.set_size_inches(15.5, 5.5)

plt.title('Cover song similarity distance: %s' % distance)
plt.xlabel('Yesterday accapella cover')
plt.ylabel('Yesterday - The Beatles')
plt.imshow(score_matrix, origin='lower')

print('Cover song similarity distance: %s' % distance)

## Computing cover song similarity distance between Yesterday - accapella cover and Come Together cover - The Aerosmith.

score_matrix, distance = estd.CoverSongSimilarity(disOnset=0.5,
                                                  disExtension=0.5,
                                                  alignmentType='serra09',
                                                  distanceType='asymmetric')(false_pair_crp)

fig = plt.gcf()
fig.set_size_inches(15.5, 5.5)

plt.title('Cover song similarity distance: %s' % distance)
plt.xlabel('Yesterday accapella cover')
plt.ylabel('Come together cover - Aerosmith')
plt.imshow(score_matrix, origin='lower')

print('Cover song similarity distance: %s' % distance)