Python NMF/HPSS Results
For this part of the project, librosa was used to analyze the bass and drums of an audio file called Music Delta - Disco. The individual bass and drums files were added together, and spectrograms and Mel spectrograms were created. The spectrogram, Mel spectrogram, original bass file, and original drums file are shown below:​​​​​​
Different separation techniques were performed on each of the spectrograms.
NMF
In performing NMF, we used the following parameters:
​
-
n_components: 12
-
init: nndsvda
-
solver: mu
-
beta_loss: kullback-leiber
-
max_iter: 500
-
random_state: 0
-
l1_ratio: 0.1
​
After reconstruction, we obtained the following spectrograms:​​​
We obtained the following audio files:
NMF seemed to separate drums very well from the audio. However, we noticed that there were noticeable drum components present in the bass output. The bass output was capturing some of the lowest and highest frequency drum sounds.​
HPSS
In performing HPSS, we used a power of 1.5 and also used soft masks.​​​ After reconstruction, we obtained the following spectrograms:​​​
We obtained the following audio files:
HPSS seemed to also separate drums well. However, it was better than NMF at attributing the lower frequency drum components to the drum output correctly. The higher frequency snares were still present in the bass output.
HPSS and NMF
To attempt at improving our results, we first performed HPSS and then performed NMF on its output.​​​ After reconstruction, we obtained the following spectrograms:​​​
We obtained the following audio files:
The bass and drums output using regular spectrograms sounded very well separated. However, the bass output was very quiet, even with a gain applied. Additionally, the Mel spectrogram output had significant distortion in its sound.
Spleeter
Spleeter was used to separate the mixed audio as well. We obtained the following audio files:
As expected, Spleeter's bass and drums outputs very accurately represented the original bass and drums input. We noticed that the bass output was also quieter than the original bass input, which indicates that bass sounds are generally very difficult to separate.