Wow, I find it incredible that this works. As I understand it, the approach is to do a Fourier transform on a couple seconds of the song to create a 128x128 pixel spectrogram. Each horizontal pixel represents a 20 ms slice in time, and each vertical pixel represents 1/128 of the frequency domain.
Then treating these spectrograms as images, train a neural net to classify them using pre-labelled samples. Then take samples from the unknown songs, and let it classify them. I find it incredible that 2.5 seconds of sound represented as a tiny picture captures information enough for reliable classification, but apparently it does!
One reason might be that the mentioned genres are highly formulaic to begin with. The standard rap song contains about 2 bars of unique music stretched out over 3 minutes with slight variations. Same with dubstep and techno. All highly repetitive. Classical music got no drums, so you can detect that. Metal got guitar distortion all over the spectrum. So with these examples the spectral images should have enough distinctive features that can be learned. Why should it be different than with 'normal' pictures? Also it looks like they take four 128x128 guesses per song.
It's true that having very different genres helps the model a lot. It would be much more difficult to distinguish between closer genres, especially when people don't really know which is which and argue all the time about it.
It's quite possible that it's mainly using even more surface-level audio features, before getting to whether the genres are formulaic or not. For example, if specific mastering studios have telltale production features visible in the audio (choice of dynamic range compression algorithms, mixing approaches, etc.), and some mastering studios mainly master, say, country, you can learn to classify country with pretty high accuracy by just recognizing a half-dozen studios' production signatures, without learning anything fundamental about the genre. Whether this happens depends a lot on your choice of data set and validation method.
From the description in the walkthrough, it doesn't. The final output looks to be based on 5 of these slices, with each providing a probability distribution that ultimately influences the final classification.
I wonder if training another net on top of the slices would work better than voting for a single winner. I'd presume that there are genres that are well characterized by the distribution and progression of their spectrograms. Probably expand/compress the collection of slices to a standard length before training?
(Nice to see you show up for the discussion. I was worried that you'd given up hope before your article hit the front page.)
To the author: Have you tried to use a logarithmic frequency scale in the spectrogram? [1] That representation is closer to the way humans perceive sound, and gives you finer resolution in the lower frequencies. [2] If you want to make your representation even closer to the human's perception, take a look at Google's CARFAC research. [3] Basically, they model the ear. I've prepared a Python utility for converting sound to Neural Activity Pattern (resembles a spectrogram when you plot it) here: https://github.com/iver56/carfac/tree/master/util
I don't think this problem is bound by absolute frequency resolution, the tightest distance between two notes on a typical piano is ~2hz and if you assume a doubling between octaves you're at <90 notes. The temporal changes and relative chord progressions probably give more info.
Thanks for your insights! I agree that log/mel spectrograms could be even more detailed and effective, and could be used with the SoX patch discussed here https://sourceforge.net/p/sox/feature-requests/176/.
That's pretty cool, I'd like to use something like this to tell me what genre my own songs are. It's annoying to write a song and then upload it to some service or another and have no idea what genre to pick. :-) My stuff is somewhere in the jazz-influenced singer-songwriter american piano pop realm which is a combination that works for me but it generally feels like I'm selling the song short if I have to pick only one.
Unless I'm misunderstanding the validation set, I'm skeptical of the ability of this classifier to tag unlabeled tracks, given that it is only being trained and tested on tracks which are already known to belong to one of the few trained genres. I'd be curious to see the performance if you were to additionally test on tracks which are not any of (Hardcore, Dubstep, Electro, Classical, Soundtrack and Rap), with a correct prediction being no tag.
Hmm, convolution is perfectly good operation to run on wave forms as well. In fact the wikipedia article (https://en.wikipedia.org/wiki/Convolution) shows the operation on functions which would correspond to time-domain wave forms. What is the point of converting everything to pictures and then using 2D convolutions when that step could have been skipped entirely?
Converting to pictures is unnecessary. It makes the processing harder. The pooling should just happen on segments of the wave form instead of the fourier transform (frequency-domain) picture spectrograms.
The idea is that the vertical axis of the spectrogram is basically already an hierarchical set of features (in scale/frequency). Then convolutions on that is a lot like how DenseNets combine hierarchical features.
I agree it seems a little jank, but the features are pretty good - and a lot of network architectures / training techniques are most practiced in an image processing context.
Thanks for your inputs, it's true that we can use convolutions on raw waveform, however the main reason I've used a spectrogram was to work on precomputed relevant features as highd pointed out, instead of running the convolution on lots of data.
1. I wonder how the continuous wavelet transform would compare to the windowed Fourier transform used here. See [1] an python implementation, for example.
2. The size of frequency analysis blocks seems arbitrary. I wonder if there is a "natural" block size based on a song's tempo, say 1 bar. This would of course require a priori tempo knowledge or a run-time estimate.
i'm not super familiar with deep learning so forgive me if i'm missing some nuance, but what's the purpose of writing/reading to/from images? seems like it would add a ton of processing time. could the CNN not just read from a 50 item array of tuples representing the data from the 20ms slice?
I'm not sure what you mean, but I have chosen to store slices on the disk so that I could still take a look at them, and not store the data only in numpy arrays. That could be optimize for a better processing time!
Of course those are going in the other direction, not generating the classification from the data, but it's probably one of the best data sets as far as classifying existing music.
Then treating these spectrograms as images, train a neural net to classify them using pre-labelled samples. Then take samples from the unknown songs, and let it classify them. I find it incredible that 2.5 seconds of sound represented as a tiny picture captures information enough for reliable classification, but apparently it does!
reply