Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Building a Music Recommender with Deep Learning

mattmurray.net

Post reply on HN
Print view View original

61–70 of 74 posts

Re: Building a Music Recommender with Deep Learning

#61

Post by Drdrdrq »

cubano wrote: post #16
madmax108 wrote: post #5

This is interesting, but fairly easy to confuse. Esp. would be interesting to see what results come up when you use modified "artistic" spectographs like that of Windowlicker by Aphex Twin [1]. One thing I've learned from years of having worked with audio and images is that image representations of audio are horrible representations of it (other than for temporal changes). The results are good though! Good work! :D […

> ...image representations of audio are horrible representations of it (other than for temporal changes). Yes, and thus the reason why the classifier was so good at recognizing trance...it's one of the few genres that locks in at around 144bpm.

What would be a better representation?

Re: Building a Music Recommender with Deep Learning

#62

Post by Nimitz14 »

johnlbevan2 wrote: post #12

Warning: this comment has little to do with the article, beyond being a rant on the approach taken by all recommendation engines I've seen. This an interesting approach, but the objective is similar to most recommendation engines: "Find me something similar to something I like". Sometimes that's a good requirement (e.g. when trying to queue up the next song in a playlist, it's good to have some similarity to the song…

Yup. Spotify is terrible at this, I religiously listen to release radar and your music of the week and the hitrate is probably 1/100?

Re: Building a Music Recommender with Deep Learning

#63

Post by JKCalhoun »

amelius wrote: post #8

> Wouldn’t it be cool if you could discover music that was released a few years ago that sounds similar to a new song that you like? Perhaps. But of course, this is likely to put the user literally into an "echo chamber" :)

Yeah, sorry, old-timer here. I loathe "genres" and strip them off my purchased music.

Is R.E.M. "Alternative", "Rock", "College"? Maybe you consider an album like "Reckoning" from R.E.M. "Rock" but then it includes a track like "Rockville" that is perhaps "Country"?

Genre makes sense for "Soundtrack" or perhaps "Classical"? But beyond that it's just mental gymnastics.

And given how fondness for music is qualitative, I've always been suspect of any sort of algorithm that tries to recommend music based on fast-Fourier-transforms. Maybe AI isn't for everything....

Re: Building a Music Recommender with Deep Learning

#64

Post by ssalazar »

Matumio wrote: post #56
personjerry wrote: post #48

My understanding of convolutions is that it's a way of extracting patterns from images. To convert audio into an image and then create convolutions from that seems... convoluted, if you will. I imagine a better way would be to think of what the equivalent of a convolution would be in the audio space? I.e. noise detection, treble/bass filters, etc.?

Convolution is generic signal-processing. It's quite common to use a one-dimensional convolution for audio filters, it would work perfectly fine as a bass filter for example. However, 2D conv+maxpool is an image processing technique that gets you translation invariance. Fine for the time dimension of the spectrogram, but rather dubious for the frequency axis. Surely you'd want to distinguish if some feature happens a…

> Fine for the time dimension of the spectrogram, but rather dubious for the frequency axis.

MFCCs[1] are exactly that, a type of convolution along the frequency axis of a Fourier transform, and are highly apt features for music classification tasks.

It makes sense if you think of timbre as a time-varying relationship between the harmonics of a single pitch; translation invariance along the frequency axis can tell you that you there are partials typical e.g. of a guitar or of a flute, without caring what particular pitch those instruments are playing. And timbre is a bigger source of variety in popular music than e.g. the particular notes used.

[1] https://en.wikipedia.org/wiki/Mel-frequency_cepstrum

Re: Building a Music Recommender with Deep Learning

#65

Post by ssalazar »

madmax108 wrote: post #5

This is interesting, but fairly easy to confuse. Esp. would be interesting to see what results come up when you use modified "artistic" spectographs like that of Windowlicker by Aphex Twin [1]. One thing I've learned from years of having worked with audio and images is that image representations of audio are horrible representations of it (other than for temporal changes). The results are good though! Good work! :D […

> image representations of audio are horrible representations of it

The spectrogram is just a series of FFTs taken over time; encoding it as a bitmap doesn't really change this, aside from precision issues. Any other representation of the audio is derived from either the original time-domain signal or the FFT.

Indeed, humans can't reliably map raw waveforms or spectrograms to intuitive musical phenomena. But a CNN should be able to derive meaningful features from these basic representations on its own.

Re: Building a Music Recommender with Deep Learning

#66

Post by 1r3n1cus »

mattmcknight wrote: post #23
oscii wrote: post #18

In my opinion, the results are not quite exciting as they might seem like at the first glance. The hip-hop and minimal house classification perform almost randomly (the random classifier would have accuracy of 50%). The claim of music genre subjectivity is not fully appropriate for the categories used in this work: the presented genres are quite distinct, and they have objective differences. Knowing only BMP and rhyt…

"The hip-hop and minimal house classification perform almost randomly (the random classifier would have accuracy of 50%). " You are assuming that this is a series of binary classifiers. It is multiclass classification, so the base rate for nine classes is 11%.

If the classes are balanced, that is. Without knowing the distribution of the classes it is difficult to understand if the result is good or not.

Re: Building a Music Recommender with Deep Learning

#67

Post by unityByFreedom »

LeCun recently griped about this topic w/rt classical music on GooglePlay,

https://www.facebook.com/photo.php?fbid=10154605399547143

> Don't you guys realize that putting everything from Monteverdi to Bach, Mozart, Beethoven, Brahms, Moussorgsky, Stravinsky, and Bernstein in the same "Classical" bucket makes no sense?

> (Particularly when you have ultra fine-grained categories for popular music!)

Any comments about that?

Re: Building a Music Recommender with Deep Learning

#68

Post by redsparrow »

1r3n1cus wrote: post #66
mattmcknight wrote: post #23

Earlier quoted context omitted.

"The hip-hop and minimal house classification perform almost randomly (the random classifier would have accuracy of 50%). " You are assuming that this is a series of binary classifiers. It is multiclass classification, so the base rate for nine classes is 11%.

If the classes are balanced, that is. Without knowing the distribution of the classes it is difficult to understand if the result is good or not.

The author downloaded 1000 tracks from each class, so they are evenly distributed.

Re: Building a Music Recommender with Deep Learning

#69

Post by johnlbevan2 »

flashman wrote: post #26
johnlbevan2 wrote: post #12

Warning: this comment has little to do with the article, beyond being a rant on the approach taken by all recommendation engines I've seen. This an interesting approach, but the objective is similar to most recommendation engines: "Find me something similar to something I like". Sometimes that's a good requirement (e.g. when trying to queue up the next song in a playlist, it's good to have some similarity to the song…

> Popular songs (i.e. if almost everyone has a Beetles album in their playlist, getting "people who bought this also bought" recommendations for anything would list Beetles I've been learning recommendation engines by looking at peoples' Steam games libraries. One feature of the data set is that many, many people own multiple versions of Counter-Strike as well as Team Fortress 2. So "a high number people who bought […

Nice approach; great to see others thinking about the issue (and unlike me, actually developing something that does something about it).

Re: Building a Music Recommender with Deep Learning

#70

Post by johnlbevan2 »

bsenftner wrote: post #30
johnlbevan2 wrote: post #12

Warning: this comment has little to do with the article, beyond being a rant on the approach taken by all recommendation engines I've seen. This an interesting approach, but the objective is similar to most recommendation engines: "Find me something similar to something I like". Sometimes that's a good requirement (e.g. when trying to queue up the next song in a playlist, it's good to have some similarity to the song…

My problem with all music recommendation engines, and for many intellectual music aficionados, the lyrics content - what is being verbally described in the music - is what I seek and hang on for my preferred music. When I listen to my collection, the genres are all over and I don't even know them. I listen to the words and treat the music as emphasis for the words. I'll have ska, 30's jazz, hip hip, and classic rock…

Not something I'd ever thought of (I tend to tune out the words / mostly treat them as another instrument; unless listening to something especially witty).

Great suggestion / I guess this leads to the idea of needing a meta recommendation engine; i.e. some way to decide what recommendation engine best works for you; selecting from one that follows lyrical themes, another that discovers "out there" content, one for similar content, etc.

Post reply on HN