User Tools
**This is an old revision of the document!** ----
====== Can we break it? Yes we can! ====== | **Authors **| Patricio López-Serrano, Hendrik Schreiber, Jonathan Driedger, Richard Vogl, Sebastian Böck| | **eMail** | [[patricio.lopez.serrano@audiolabs-erlangen.de]],[[jonathan.driedger@audiolabs-erlangen.de]],[[hs@tagtraum.com]]| EDM (Electronic Dance Music) is often constructed from //samples// or //loops//, stacked on top of each other. A common pattern of EDM artists is to start a song with a single loop. While the song builds up, more and more loops are added. {{ :step1.jpg?nolink&400 |}} As an example, listen to this excerpt of Moby's song "Porcelain". {{ :Moby_Porcelain_FullSong.mp3 |}} The goal of our hack was the following: Given an EDM song that was built up in the way described above, we want to **break it into all the separate loops that were used to create it**. To this end, we came up with an iterative strategy. ===== Step 1: Identifying Isolated Loops ===== In a first step, we need to identify a loop that occurs without other loops stacked on top of it. {{ :step2.jpg?nolink&400 |}} In our example, this is this loop {{ :Moby_Porcelain_Iteration1_Loop.mp3 |}} We achieve this by first computing a //self-similarity matrix// (SSM) of a feature representation of the song and identifying a segment in the recording that is frequently repeated. To account for the fact that the repetitions of the sample are most probably superimposed with other loops, we experimented with feature representations and distance measures which take this circumstance into account. ===== Step 2: Find the Reoccurrences of the Identified Loop ===== Now, we need to identify the positions in the song where the identified loop is also present. {{ :step3.jpg?nolink&400 |}} We do this by computing a //matching curve// and applying a //peak picking// scheme. Again, suitable feature representations and distance measures played an important role. ===== Step 3: Remove the Loop from the Song ===== Now, we remove the loop from the song at the identified time positions. This is done in the time-frequency domain where we subtract the spectrogram of the loop point-wise from the spectrogram of the full song (we subtract the magnitudes, capping at 0, and use the phase information of the original song). {{ :step4.jpg?nolink&400 |}} The //remainder// of the song in our example sounds like this {{ :Moby_Porcelain_Iteration1_Remainder.mp3 |}} ===== Step 4: Iterate ===== To remove the next loop, we start from the beginning again. {{ :step5.jpg?nolink&400 |}} In our example, the next loop to remove is {{ :Moby_Porcelain_Iteration2_Loop.mp3 |}} ===== Results ===== Here, you find the audio examples that we got from performing 3 Iterations of our procedure |Full song |{{:Moby_Porcelain_FullSong.mp3|}}| Iteration 1 : SynthPad |Loop|{{:Moby_Porcelain_Iteration1_Loop.mp3|}}| |Remainder|{{:Moby_Porcelain_Iteration1_Remainder.mp3|}}| Iteration 2 : Drums - soft |Loop|{{:Moby_Porcelain_Iteration2_Loop.mp3|}}| |Remainder|{{:Moby_Porcelain_Iteration2_Remainder.mp3|}}| Iteration 2 : Drums - strong |Loop|{{:Moby_Porcelain_Iteration3_Loop.mp3|}}| |Remainder|{{:Moby_Porcelain_Iteration3_Remainder.mp3|}}| I.e. the task is: - Find a sample that's as //pure// as possible. - Removethe sample wherever it occurs in the song. - Repeat. //Pure// in this context is shall mean //standalone//, i.e. a sample that is not polluted by another sample yet. ===== Finding pure Samples ===== - SSM (diagonal median filtered) - Identify paths that have a certain length and length/distance relationship - Pick a path according to the super-secret rules - Find similar locations in the song To find similar locations in the song, we weren't able to use a regular cosine distance function, because we are not looking for absolute similarity, but rather a "Sample A contains Sample B" relationship. ===== Removing samples from the Song ===== In order to remove samples (loops) from the song, we simply subtract the magnitude spectrogram from the original spectrogram and re-synthesize both signal—the remains of the song and the sample we subtracted. ===== Repeat ===== We can continue this process for a number of iterations. Results get stranger and stranger...
