I’ve always wanted to learn how to separate vocals from a track programmatically and not depend on software-as-a-service to perform the task for me. This article illustrates how to separate the vocals of a song from the instruments using my new favorite library, Librosa. You can check out the Google Colab Notebook here.
The idea sparked when I wanted to separate individual tracks of a song I really liked, so I went to Product Hunt and discovered melody ml. This discovery started the urge to learn ML for music, and the eventual discovery of the Python library, librosa.
By the way, I ran out of RAM, which made my notebook explode.
https://twitter.com/tonypoppinss/status/1620369862630739968?embedable=true
Install and import dependencies
pip install librosa matplotlib IPython
import librosa
from librosa import display
import numpy as np
import IPython.display as ipd
import matplotlib as plt
Load and display the song.
I used My Last Serenade by KSE as I wondered how the growling or shouting parts of the song would come out.
y, sr = librosa.load('My Last Serenade.wav')
ipd.Audio(data=y[90*sr:110*sr], rate=sr)
We slice a 20-second snippet in the chorus of the song. We show the audio using ipd.Audio(tbh, this is a bit exhausting). The photo is shown below because I couldn't find a way to upload audio here on DEV.
We separate a complex-valued spectrogram D into its magnitude (S) and phase (P) components, convert the time stamps into frames, plot the data, then display the full spectrogram of the data.
S_full, phase = librosa.magphase(librosa.stft(y))
idx = slice(*librosa.time_to_frames([90*110], sr=sr))
fig, ax = plt.pyplot.subplots()
img = display.specshow(librosa.amplitude_to_db(S_full[:, idx], ref=np.max), y_axis='log', x_axis='time', sr=sr, ax=ax)
fig.colorbar(img, ax=ax)
Line-by-line explanation
S_full, phase = librosa.magphase(librosa.stft(y)) - we separate the magnitude and phase of the track using short-time Fourier transform by representing a signal in the time-frequency domain by computing discrete Fourier Transforms(DFT)(y)
idx = slice(*librosa.time_to_frames([90*110], sr=sr)) - slice the part of the song then convert it to stft frames using the time_to_frames function of librosa.
img = display.specshow(librosa.amplitude_to_db(S_full[:, idx], ref=np.max), y_axis='log', x_axis='time', sr=sr, ax=ax) - display the spectrogram of the 20-second sliced part of the song by converting the amplitude spectrogram to a dB-scaled spectrogram of the magnitude, then compares the magnitude and phase of the track and returns a new array containing the element-wise maxima and finally it plots the y and x-axes.
Below is the image of the spectrum:
Decomposing the spectrogram
S_filter = librosa.decompose.nn_filter(S_full, aggregate=np.median, metric='cosine', width=int(librosa.time_to_frames(2, sr=sr)))
S_filter = np.minimum(S_full, S_filter)
Line-by-line explanation
S_filter = librosa.decompose.nn_filter(S_full, aggregate=np.median, metric='cosine', width=int(librosa.time_to_frames(2, sr=sr))) - we filter the vocals by their nearest neighbors, aggregate their median values, compare their frames using cosine similarity and contain those frames to be separated by 2 seconds and suppress other sounds from the spectrum
S_filter = np.minimum(S_full, S_filter) - we get the calculated data in the memory of the S_full and S_filter variables to get the minimum value.
Display the background and foreground spectrum of the audio
margin_i, margin_v = 3, 11
power = 3
mask_i = librosa.util.softmask(S_filter, margin_i * (S_full - S_filter), power=power)
mask_v = librosa.util.softmask(S_full - S_filter, margin_v * S_filter, power=power)
S_foreground = mask_v * S_full
S_background = mask_i * S_full
Line-by-line explanation
margin_i, margin_v = 3, 11 - we use margins to reduce loss in sound in the vocals and instrumented masks
power = 3 - returns the soft mask computed in a numerically stable way
S_foreground = mask_v * S_full and S_background = mask_i * S_full - multiply the masks with the input spectrum to separate the components
Plotting the full spectrum, background and foreground spectrum
fig, ax = plt.pyplot.subplots(nrows=3, sharex=True, sharey=True)
img = display.specshow(librosa.amplitude_to_db(S_full[:, idx], ref=np.max), y_axis='log', x_axis='time', sr=sr, ax=ax[0])
ax[0].set(title='Full Spectrum')
ax[0].label_outer()
display.specshow(librosa.amplitude_to_db(S_background[:, idx], ref=np.max), y_axis='log', x_axis='time', sr=sr, ax=ax[1])
ax[1].set(title='Background Spectrum')
ax[1].label_outer()
display.specshow(librosa.amplitude_to_db(S_foreground[:, idx], ref=np.max), y_axis='log', x_axis='time', sr=sr, ax=ax[2])
ax[2].set(title='Foreground Spectrum')
ax[2].label_outer()
fig.colorbar(img, ax=ax)
Recover the foreground audio from the masked spectrogram and playback the audio
y_foreground = librosa.istft(S_foreground * phase)
ipd.Audio(data=y_foreground[90*sr:110*sr], rate=sr)
Line-by-line explanation
y_foreground = librosa.istft(S_foreground * phase) - inverses the short-time fourier transform ipd.Audio(data=y_foreground[90*sr:110*sr], rate=sr) - plays back the vocals from the track
Conclusion
This seemed easy at first thought and when I was reading the documentation but digging under the code made me realize that this idea was a little more complex. But, what made me continue was when I read about nearest neighbors in one part of the documentation which made me realize that I will be getting my hands dirty with Machine Learning in the future with this library.
Also published here.