A kick can sound enormous on its own and disappear into a mess once the bass arrives. A textured snare can be fascinating until its long tail covers the next hit. Fill sixteen pads with individually impressive sounds and you still have another job to do: make a kit.
That is the musical problem behind our Text-to-Kit research. We are investigating how a description of a sound world can become a collection you can arrange and play.
Similar enough to belong, different enough to arrange
Imagine a dry, close kit. Its short kick gives the bass room, a crisp hat marks the pulse, and the snare supplies the backbeat. The sounds have a family resemblance, but each one contributes something different.
Now imagine every pad has the same thick attack and lingering low end. The collection might share a strong character, yet adding a pad mostly adds more of the same weight. A coherent kit needs contrast within that character.
This is familiar territory when you build a kit from samples. You listen to candidates together, replace the hit that crowds the groove, and decide whether a strange sound deserves a separate accent. A generator faces the same musical demands. The question is how much of that relationship it can help establish before you start arranging.
What our early experiments suggested
We explored two approaches: generate individual pads under a shared style description, or generate a long recording and cut it into hits.
In the sets we studied, the separate-pad approach produced more distinct sounds and cleaner tails. That is an early observation; a matched comparison and listening study are still needed before we can say whether the advantage holds more generally.
The observation matters because a clean ending leaves you room to decide how sounds overlap. A hit can still be long or noisy when that serves the kit. An abrupt cut or an unwanted fragment of another hit gives you repair work before you can make that choice.
Likewise, distinct pads give the pattern more ways to change. Moving an accent from one pad to another should offer a different color or role, rather than a barely noticeable substitute.
Listen to the relationships
When you explore a generated kit in the Text-to-Drumkit workbench, start with a sparse pattern. Give the kick, snare, and hats room to reveal how they fit. Then add one accent sound and listen to what it contributes.
If a pad is awkward, consider whether its level, tail, or role is the problem before replacing it. A sound that crowds the backbeat might be interesting as an occasional fill. Our Audio generation guide looks at the choices between receiving a generated clip and turning it into musical material.
For our next research steps, we want comparisons that tell us more about those relationships: whether pads are meaningfully different, whether their endings are usable, and what musicians hear when they play them together. Listening to isolated hits can only answer part of the question.
Taking the questions to ISMIR and ADC
Our Text-to-Kit work has been accepted for the Late-Breaking/Demo session at ISMIR 2026, November 8–12 in Abu Dhabi.
Alongside this research, our virtual poster at the Audio Developer Conference, November 9–11 in Bristol and online, examines agent-assisted plugin creation. It asks how people can develop tools around their own musical ideas and practices.
These are opportunities to discuss the work with people who study music and make instruments. If you have experience building kits, evaluating generated sounds, or developing musical tools, tell us what you would want to hear or change. Those questions can help shape what we investigate next.
