How Super scaled real-time voice agents for real estate

AssemblyAI is one of the leading speech-to-text providers in the market. They recently interviewed Super for a deep dive into how we approach building real-time voice agents for property management and real estate use cases.

A few key learnings and insights we shared:

• Speech-to-text is the first domino. In a real-time phone conversation, speech-to-text (STT) quality determines whether the call feels like magic or like a robocall. Awkward pauses, mistimed barge-ins, or a misheard addresses erode caller trust within seconds, so we treat STT selection as a strategic engineering decision at the leadership level rather than a commodity infrastructure choice.
  ‍
• Vertical voice AI demands specialized criteria. Generic STT benchmarks don't capture what matters for property management calls. We evaluate providers on smart turn detection (well beyond basic min/max timing controls), short-utterance handling, multilingual detection, and keyterm prompting that actually boosts the property names, addresses, and contact details that come up constantly on every call.
  ‍
• Focus your engineering on what differentiates you. We invest our engineering resources in Super's multi-agent platform and orchestration layer, and partner with best-in-class providers for foundational components. That means we're constantly evaluating the best models for our specific needs so that our engineering team can focus on new product development.
  ‍
• Voice QA is its own discipline. Every model change at Super runs through scaled evaluation on our voice QA platform plus targeted human QA on the failure areas we care most about: sound quality, background noise, short replies, specific keyterms, and language detection. Depending on the depth of refinement, that's a few days to a couple of weeks of iteration before anything reaches production.
  ‍

Read the full interview on AssemblyAI's site.

‍