AWS Startups Blog

Splice Leverages ML to Connect Musicians With the Building Blocks for Creativity

Splice Header Image

Over the years, the physical act of producing music has become increasingly technical, which is both a blessing and a curse for musicians of all levels. On one hand, artists have access to more tools of the trade than ever before. On the other, the sheer volume of possibilities is enough to stop someone in their tracks. Creative blocks are real, and there’s no quicker path to artistic purgatory than obsessively searching for the sound in your head for hours and coming up empty.

That’s where Splice comes in, a creative platform for musicians, built by musicians, to empower artists to unleash their true creative potential. One of its core offerings is Splice Sounds, the largest royalty-free audio sample library in the world, giving users access to millions of sounds and presets for a monthly fee of $7.99. “As our catalog grows, so does the challenge of finding the right sound. That’s why we are investing in building best-in-class search and discovery capabilities.” says Alejandro Koretzky, Head of Machine Learning & Principal Engineer at Splice.

Alejandro Koretzky, Head of Machine Learning & Principal Engineer

Alejandro Koretzky, Head of Machine Learning & Principal Engineer

The company was founded in 2013 and now caters to more than 3 million musicians that explore the catalog in search of the perfect sounds. As if having the largest library of samples and sounds wasn’t enough, Splice also curates artist packs, which contain original sounds from the world’s most talented producers.

Another large barrier to entry in the production space is the cost of software and plugins, which can easily stack up. To ease the financial burden and prevent piracy, Splice introduced a rent-to-own model, which allows users to try a plugin for 3 days, then pay a small monthly fee until they own the plugin outright.

But back to a core problem of modern musical production: When you have a certain sound in your head, but you can’t quite articulate it through known search terms, what do you do? Splice’s answer: Let machine learning take the wheel. Enter ‘Similar Sounds,’ a newer user-facing offering from Splice that aims to make it easier than ever to connect musicians with the sounds they’re looking for.

“Compared to images, sound is a lot harder to describe using words, especially when you’re looking for something very specific. Similar Sounds is the perfect complement to text-based search and it’s allowing our users to discover and navigate our catalog in ways that weren’t possible before”, says Alejandro.

Although the new feature is still in relative infancy compared to other Splice products or initiatives, it’s already showing massive signs of success. In fact, the company has seen a near 10 percent increase in search conversions since its launch.

Hannah Mimi Park, Tech Lead in the Search and Recommendations team, is very optimistic about this hybrid approach to search: “searching by sound is a feature I’ve wanted to build for a long time, and something I think about myself as a songwriter: there is no set search lexicon for music search, and while text-based search is core to almost every content platform including ours, when it comes to music, it is just not enough.”

Delivering similarity search across the largest audio sample library in the world was not without challenges. It involved training several ML models from scratch, and building complex infrastructure and pipelines for fast inference and content retrieval. The team at Splice relied on many of AWS’s cloud services to make this happen.

“We’ve been using Amazon SageMaker as a core part of our data preprocessing, training and inference pipelines. Although much of our system is built custom as a result of the types of data we work with, our team has been really happy in leveraging SageMaker as a launchpad to run our training jobs and for deploying API endpoints”, says Naveen Sasalu Rajashekharappa, Senior ML Engineer at Splice.

Similar Sounds is also being used by artists to discover new content that they may have glazed over or missed given the sheer size of Splice’s content library. This includes sample packs that aren’t as popular amongst the membership. It may have been introduced as a quicker or more efficient way to search and surface the right sounds, but it’s also connecting musicians with new possibilities that they might not have discovered on their own, which is the core of Splice’s mission.