"I we didn't find Resemble AI, I do not know if we could have been able to make the game."
- Karleigh Sheen-Drumm, Producer, RED Games
.avif)
Meet Karleigh
Karleigh Sheehan-Drumm is a producer at RED Games. She has no acting background. No voice training. She describes herself as “a very confident reader.”
Two years into building Crayola Adventures, her voice model had more training data than almost anything else RED Games had ever produced.
She can now pick her own voice out of a lineup regardless of what speech model it has been run through. She has heard herself speak German, French, Japanese, and 10 other languages she does not speak.
None of that was the plan. Somebody had to do it, and so it fell upon her.
THE PROBLEM
The game nobody knew how to build
Crayola Adventures is a Mad Lib-style storybook game for early readers on Apple Arcade. Players type in or speak any word they want. The game reads it back, mid-sentence, in a live narrated performance. Your favorite color, your favorite noun, your favorite kind of music, all of it gets woven into the story as it unfolds.
The concept was creatively ambitious. It was also technically unsolvable out of the box.
The team knew early on that narration was non-negotiable. They had built children’s titles before. They knew younger players needed to hear the story, not just read it. The game only works if a kid types “spaghetti” and hears it read back, mid-sentence, in the exact same voice as every scripted line around it.
Three options were on the table:
- Record a fixed library of nouns and adjectives and restrict what players could type.
- Ship with plain TTS and accept that players would notice it wasn’t real.
- Pivot the entire game concept.
There was also a hard technical requirement that made all three options worse: the game had to launch in 13 languages for Apple Arcade. Whatever the solution was, it had to scale across all of them.
"Fundamentally, the whole experience would have been a lot more restricted without Resemble AI. The end result would have not been as compelling. "
- John Karczewski, Engineer, RED Games
THE SOLUTION
The pipeline they built
RED Games found Resemble AI. What followed was not an off-the-shelf integration. It was a custom pipeline built in close collaboration with Resemble’s team, combining three distinct capabilities into a single production workflow.
Karleigh recorded every scripted line as the base voice performance. Resemble’s Speech-to-Speech mapped that performance onto each character in the game’s full cast. For the Mad Lib blanks, a separate TTS model, trained on Karleigh’s voice, generated player-supplied words on demand. Those words passed through a second Speech-to-Speech layer to match the target character’s voice and the cadence of the line around it.
Every word a player typed got read back at the same quality as the lines Karleigh recorded in a studio.
"It was a full ecosystem of my initial read being mapped onto a new voice. My voice data was also informing the words themselves from the text to the speech."
- Karleigh Sheen-Drumm, Producer, RED Games
Resemble set RED Games up on a dedicated endpoint, scaled to handle Apple Arcade launch volume. The team pre-batched 10 to 20 upcoming narration lines so synthesis completed before a player reached the next story beat. Average latency per clip stayed around 1 to 2 seconds. Total batch turnaround held under 30 seconds.
"Resemble ended up merging features into a specific pipeline specific to our use case. It always felt really like a partnership."
- John Karczewski, Engineer, RED Games
THE AUDIENCE IMPACT
What a six-year-old actually experiences
A child sits down with Crayola Adventures. The game asks for a favorite food. They type “spaghetti.” A few seconds later, the narrator, in full character voice, reads it back mid-sentence. The dragon says “spaghetti.” The story keeps moving. The child laughs.
That moment is the product. Everything above, the pipeline, the latency targets, the two-stage Speech-to-Speech architecture, exists to make that moment feel completely normal to a six-year-old.
THE RESULTS
What it produced
Crayola Adventures won an Apple Design Award. Apple gives fewer than 15 of these per year across all app categories. Voice was central to the win.
The game has been live on Apple Arcade since launch, accumulating 1,000,000 downloads and installs. RED Games monitors synthesis times in production. The pipeline has stayed within range for over two years.
Resemble generated 26,000,000 unique lines of dialogue for Crayola Adventures alone since launch, representing 22,000 hours of synthesized audio.
"The speed to be able to turn stuff around in the dashboard has made life so different."
- Karleigh Sheen-Drumm, Producer, RED Games
WHAT THEY BUILT NEXT
One pipeline. Three titles. Still running.
RED Games did not stop at one game. The same Resemble pipeline now runs across two additional titles: Scribble Scrubbie Pets and Crayola Create and Play. Karleigh’s voice model has been extended to each of them.
The marketing team at RED Games now uses the Resemble dashboard independently, without engineering support, to turn around ad reads and cross-promotional content on short notice.
One voice actor. One pipeline. Three titles. Thirteen languages. Still running.
PRODUCTS USED
Speech-to-Speech / Text-to-Speech / Localization
