Back in 2011, on a spring edition of Los Angeles' The Do-Over block party, headliner DJ Jazzy Jeff started playing Herbie Hancock's "Jessica"—seemingly an odd choice to get a crowd going. But as soon as he loops two bars and pitches that down by 14 semitones, the melody of "Shook Ones PT II" by Mobb Deep begins to take shape. Questlove comes rushing to the DJ booth to see it with his own eyes: right then and there, a sample mystery that had been around since 1995 was unriddled live on the spot, shortly after it was discovered by Timon 'Bronco' Heinke on music forum The Breaks. The response by Questlove is priceless.
That beautifully captured moment is just one example of how cratediggers have unearthed samples for as long as the art of sampling itself exists. That's been thoroughly documented—and widely debated—from late-80s compilations such as Ultimate Breaks And Beats to today's online databases like WhoSampled. The thought of artificial intelligence discovering a sample isn't as charming as the sight of a mind-blown Questlove geeking out together with Jazzy Jeff. But the significance is huge: the technology breaks entirely new ground in identifying samples. With a method that was discovered only recently. Covertly pioneered by members of a Discord community by the name of Sample Hunting.
"Google Assistant can even detect samples less than a second long, and is usually able to detect samples that have been chopped or time-stretched."
Not long after Daft Punk's break-up in 2021, several sampling fiends bonded online over a mutual obsession to find the unknown samples on Daft Punk's seminal 2001 album, Discovery. More specifically, in the sample-dense "Face To Face" co-produced by Todd Edwards.
The founder of the Sample Hunting community goes by the moniker of lobelia. She already dabbled with Google and Shazam to detect samples around 2016/2017. But it took another five years before she realized this could be a game-changer. "When Google Assistant helped me find 'South City Midnight Lady' by The Doobie Brothers as a guitar sample in 'Face To Face' in late 2021, I realized that this method could be huge," she recalls. "Especially because, at that point, we didn't even know that sound was a separate sample. We actually thought it was part of another sampled record."
"I slept through what we now call 'the Night of Many Samples' when I'd say a dozen samples were found. I can't describe how crazy waking up to all of that was!"
But it wasn't until mid-2022 that Google's song recognition turned from another mid-Shazam alternative to a groundbreaking discovery for them. A Sample Hunting member by the name of DJPasta found a new way to utilize the technology to the fullest: "I figured out a method to run audio directly from my PC into Google Assistant with software called Bluestacks. I was mostly trying out a few Todd Edwards samples that I'd been looking for at the time. To my surprise, Google Assistant's song recognition found most of them. Eventually, I had the idea to try out shorter samples, like Carrie Lucas' 'Sometimes a Love Goes Wrong.'"
A stint of discoveries followed, also by other members of the community who started using Google Assistant. From there, they had pretty much taken Google Assistant on board as the new default for sample hunting, on top of their continued discoveries by ear, knowledge of music, and labor of endlessly digging through music.
Lobelia recalls there was a sample drought for Daft Punk's "Face To Face" up till July 2022. Lobelia: "I slept through what we now call the Night of Many Samples when I'd say a dozen samples were found. I can't describe how crazy waking up to all of that was!" Since then, they tongue-in-cheek call Google's song recognition technology The Blessed AI. A divine status with the community members as its disciples.
That led them to unravel previously undiscovered samples in music by Mobb Deep (1996's "Hell On Earth (Front Lines)"), Madlib / Quasimoto (2000's "Green Power"), Nujabes (2004's "Decade (Interlude)" and 2010's "Another Reflection"), Daft Punk (2001's "Too Long" and numerous samples on "Face To Face"), and French house duo Modjo (2001's "Music Takes You Back"), among others. It's an ever-growing list of samples that were shrouded in mystery for over two decades. Ones that even the most seasoned diggers hadn't found before. Now, artificial intelligence is outsmarting them.
Both Shazam and Google Assistant use similar audio fingerprinting methods. But as exemplified by functionalities like Hum to Search in 2020, Google's use of deep neural networks makes the tech behind their song recognition far more advanced.
"Google Assistant can even detect samples less than a second long, and is usually able to detect samples that have been chopped or time-stretched," explains DJPasta. According to lobelia, that's far more accurate than alternatives like Shazam: "With Shazam, we usually had to try and match the tempo and structure near-perfectly to get a result. We usually go against using Shazam anyway. Because, for some reason, Shazam likes to suggest random EDM tracks from the 2010s. Not very helpful when you're looking for a jazz record..."
There are also ways to trick Google Assistant into giving successful results for short samples. DJPasta: "If the sample you are trying to find is part of a much longer drawn-out chord or texture, you can time-stretch or crossfade a loop to make the AI think it's longer. You can also repitch the sample to form a chord progression if you guess it correctly." Another useful 'hack' according to DJPasta is the use of AI stem separation, or manually extracting certain elements with toolkits such as iZotope RX.
"Considering this method is only in its infancy, this sets a precedent for the future of what's commonly known as sample spotting."
So should producers start to worry now? What's next after audio fingerprinting systems like YouTube's Content ID have already proven to be quite effective when it comes to detecting copyrighted material? The difference is that this new method puts control into the hands of anyone looking for certain samples, as opposed to relying on a fingerprinting system fully controlled by a tech giant.
"Something like this is more effective to detect much shorter and manipulated samples," says DJPasta. "If it's [a rightsholder's] own release, they also have access to multitracks which means they don't need to manually edit the audio to isolate any samples."
Considering this method is only in its infancy, this sets a precedent for the future of what's commonly known as sample spotting. "It's just a matter of time before even more people start using this particular technology as well," says Sample Hunting founder lobelia. "It gets even more reliable over time. Last year was only the beginning; we're all hungry for more."
What is Tracklib?
Tracklib is a crate-digging platform to sample and clear original music. The service makes sample clearance fast, easy, legal, and affordable for every music producer out there.
That way, producers can safely navigate the world of sampling without having to deal with clearance issues. All to be able to fully focus on what they love most: making music.
We are unfortunately unable to offer support in the comments. If you have any questions, reach out to us here!