Building Epidemic Sound’s first GenAI feature
In 2023-2024, Epidemic Sound was training data models to generate music using our proprietary music catalog as input. I, together with 2 frontend engineers and 1 UX researcher were tasked to find a way to package the technology into a product feature that would establish our company as a strong player within the AI music & soundtracking space.
The alpha prototypes
We worked in 1 week sprints where we launched a new Alpha prototype at the end of every week, which we would test with users and gather feedback that feeds into the next prototype. We kept it scrappy, lean and really focused on building to learn.
Alpha 1: First take, not much thought put into the UX yet. We wanted to connect the backend with a simple UI and get the music generation working.
Alpha 2: Added an input field for users to describe the changes they want to make + separating the output audio into stems (instruments). Learning: Users have a hard time describing what they wanted.
Alpha 3: Instead of free-text input, we tried using keywords + sliders to control input strength. Learning: Users preferred this way of providing input but the results didn’t sound good when they added “Christmas” on a hip-hop song for example. We need to set better guardrails.
Alpha 4: Evolved the keywords + sliders into pre-defined mood presets which produced better results. Learning: Users found this useful but it’s more of a “nice to have” than a feature that would solve a real pain point.
Learnings after 4 Alphas
We realized that we were focusing too much on pushing the new shiny technology and hoping that users will want it. It felt gimmicky and got reactions like "It's cool! I would play with it but won't use it in my actual work".
We took a step back to rethink how we should approach the next Alpha. What if we forget everything we know about the technology and start from users' existing pain points instead? When in the users’ workflow does it hurt the most? Can we use this new technology we have to solve it?
Hitting the nail on the head with Alpha 5
Instead of generating completely new melodies or drum tracks in a different style, what users really wanted was to be able to move parts of a track around. This is a big pain point today where they need to manually stitch parts of a track together.
We used generative AI to inpaint the seams so transitions would sound perfect. On top of that, we allowed users to easily turn off any instrument, add mood filters or change intensity at any point in the song.
Here is a click-through prototype that I made to visualize the experience.
Alpha 5 → Beta → General Availability
When we built Alpha 5, we saw a complete shift in sentiment among both the users we tested with and internally at Epidemic Sound. It wasn’t as flashy as the previous alphas, but it felt right. It was true to our journey, and more importantly, it solved a major pain point for our users.
Alpha 5 became our Beta. Together with a larger group of designers and researchers, we ran user tests on it to help us define what we should launch to the public. Once that was done, we handed over the work to another team to get the feature ready for general availability.
