Extract loops with loopextractor

December 18, 2019

I’ve just published a new resource: loopextractor, an implementation of an algorithm described in our 2018 ICASSP paper (details here) that uses non-negative Tucker decomposition to model the loops in a song.

An improved and extended version of this work was recently used to power our website Unmixer, which was presented at ISMIR last month.

Because the ICASSP paper was written while I was employed at AIST, and the Unmixer website was developed in a later collaboration, I cannot share the code that I wrote for those projects.

But, I have now rewritten the algorithm (from the 2018 paper), and I am allowed to make that public! It’s on Github as loopextractor.

As the code makes clear, the steps are very straightforward: stripping away the comments and error checks, it is only about 50 lines of code. Here’s a 9-line pseudo-code version:

# Load audio:
signal, sample_rate = librosa.load(audio_file, mono=True)
# Get downbeats from madmom:
downbeat_times = get_downbeats(signal)
# Compute spectrum for each downbeat period,
# and stack into a cube:
spec_cube = make_spectral_cube(signal, downbeat_times)
# Get Tucker decomposition with TensorLy:
core, factors = tensorly.non_negative_tucker(spec_cube,
    n_templates)
# Reconstruct each loop:
for i in range(n_loops:
    # Multiply templates to get modeled spectrum of loop:
    spectrum_i = factors[0] * factors[1] * core[:,:,i]
    # Choose the loudest instance of the loop:
    bar_i = find_loudest_bar(factors[2], spectrum_i)
    # Do soft-masking with librosa (to get phase):
    signal_i = mask_spectra(spectrum_i, spec_cube[:,:,bar_i])
    # Write audio:
    librosa.output.write_wav("output%i.wav" % i, signal_i)