Introduction to the approach:

Using the FFT approach, we “teach” a matrix to learn how to differentiate between genres. We did this by feeding the database matrix with lots of samples of each genre and tell it what genre each sample is. Then we ask the matrix to make a smart decision about a given sample’s genre.

Creating the database:

To create our database, we collected 10 songs for each genre (Rap, Rock, Jazz, Pop, and Classical). We first convert these songs from Stereo to Mono, which makes it easier for Matlab to carry. Using this we’ll get some long vectors representing each song. We divide each song into samples of length (15ms). Those short samples allow us to get a nice, identified picture of the frequencies represented by those samples. We then take the FFT (Fast Fourier Transform) of each sample, and stack them next to each other into the columns of our database matrix, with 661 rows. Each column represents the frequency spectrum of each 15ms. After that we normalize each vector respectively. We find out that Matlab would run out of memory really fast with huge amount of information. An average song of 5 minute length, for example, would have 20,000 samples of 15ms. And each one would have 661 rows, which gives us 13,220,000 numbers to represent one single song. We solved this problem by taking average of columns: we indeed average a 5 minute song into 75 columns. Although this certainly affected our algorithm’s accuracy by some degree, it allows us to add more songs to be able to build a bigger database matrix, thus compensate for the lost accuracy.

Now we label each column with a number, which represents the type of genre the column stands for (1-5). So each number would correspond to a certain FFT column.