0
$\begingroup$

For building a music recommender system(content based) which signal features should i use as feature factor ?? Will it be peridogram,spectrogram or fft or anything else ? This system will recommend music corresponding how much they are similiar. So which features correspond strongly to how much signals are similiar ? Assume correlation measures signal similiarity (but i havenot seen any implementation based on correlation on signals).

$\endgroup$
  • $\begingroup$ You can start with features calculated using some off-the-shelf audio processing library like Librosa. librosa.github.io/librosa/feature.html BTW if I knew the answer to this, I would probably have my own competing company to Pandora and Spotify :) $\endgroup$ – Atul Ingle Feb 17 '17 at 14:26
  • $\begingroup$ Thanks, I neither want to compete with them. Though i want some decent features of audio signal which can be used to build a recommender system which can give a decent result. $\endgroup$ – Nazmus Salehin Feb 17 '17 at 16:26
  • $\begingroup$ How about taking only a voice portion (not of songs, just voice). Which features can be selected for that ? $\endgroup$ – Nazmus Salehin Feb 17 '17 at 16:27
  • $\begingroup$ I would start with f0 (fundamental frequency/pitch) and various MFCC coefficients. These are easy to compute with libraries like Librosa or Chroma Toolbox. $\endgroup$ – Atul Ingle Feb 17 '17 at 16:52
0
$\begingroup$

For a big data, deep learning approach, you could try constructing an autoencoder from something such as WaveNet to achieve dimensionality reduction to get a feature vector. Alternatively, you could take STFT (short time Fourier transform) spectra of the songs over time and feed them into an RNN, again in autoencoder fashion. My intuition though is that, due to the immense complexity of musical genres, and current limitations of consumer hardware, any approach based on the raw audio alone is going to be very disappointing (unless you have a GPU farm).

Really grossly, you may get some results by finding the BPM/tempo, the spectral density, overall entropy, variance in entropy over time.

Actually, look at some of the feature extractors used in this Kaggle EEG contest. This may give you some rudimentary results, especially combined with XGB (extreme gradient boosting)

I believe most music recommendation services work like search engines, using an algorithm similar to PageRank to create a graph which associates artists.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.