AptixLabs uses ElevenLabs as its voice partner across every product that talks. The strategic bet: human-quality voice removes the single biggest friction in fitness apps, which is that text-based instructions get ignored mid-workout.
How the runtime works
For UltraFit360 we use ElevenLabs streaming voices to deliver real-time audio coaching during workouts. Text scripts written by coaches are turned into spoken audio on device-trigger, so the same coach voice can cover thousands of new workouts without re-recording. Latency-sensitive paths use the streaming endpoints; longer-form output is generated ahead of time and cached in Cloud Storage.
The brand voice clone
The studio also clones a brand voice for the automated reel pipeline so all generated content sounds the same regardless of who wrote the script. The clone is a fine-tuned variant of an ElevenLabs voice model trained on roughly 30 minutes of reference audio, with explicit consent and licensing.
How the partnership shaped the build
ElevenLabs partnered with AptixLabs early and contributed 33 million character credits, which is enough to ship lifelike audio coaching across the full UltraFit360 rollout without billing pressure. That credit pool changed what the team could prototype — features that would have been "too expensive at scale" became fair game.
What we learned
- Streaming beats pre-generation for anything triggered by a user action
- A single brand voice across products lowers cognitive load
- Caching deterministic outputs (the same prompt produces the same audio) cuts cost more than picking the cheapest voice
- Voice cloning needs explicit, written consent — get it in writing before you train
Why text-to-speech was a non-starter
Robotic voice in a fitness app is worse than silence — it breaks the one thing audio coaching is for, which is feeling like someone is in the room with you. Early on we tested platform TTS and watched users mute it within one session. Human-grade voice was not a nice-to-have; it was the difference between a feature people use and a feature people turn off. That is why we treated voice quality as a product requirement, not a cost line.
Managing the cost of speech
Generating fresh audio on every workout would be wasteful, because most coaching lines repeat. So a line spoken in one workout is cached and reused across every workout that uses it — the first user pays the generation cost, everyone after hears it instantly and for free. Only genuinely dynamic lines (a user's name, a personal record) are generated live. It is the same caching discipline we apply to every AI feature: pay once for anything deterministic.