Three generations of conversational structure
Over the previous decade, the stack behind these calls developed via three distinct eras, transferring from bespoke neural pipelines to a unified structure utilizing Gemini.
1. Pre-LLM closed-domain fashions
We began work on Google Duplex earlier than trendy language fashions and even earlier than the transformer itself was invented. In these early days, to make sure companies and customers had an excellent expertise, we constrained Duplex to slim domains and skilled customized recurrent neural community fashions on anonymized, task-specific cellphone conversations. With out pre-trained world information, we taught our fashions how you can deal with each particular sort of process effectively. In apply, this meant counting on intensive coaching information for every process sort — similar to deciphering the inflexible move of restaurant reserving versus checking vacation hours — to make sure the system might reliably navigate domain-specific logic.
2. Early generative fashions and hybrid stacks
Google Duplex began using open area basis fashions in manufacturing as they emerged. For instance, in 2022 Duplex used LaMDA in reside calls. These pre-trained fashions simplified comprehension and commonsense reasoning, permitting the system to deal with surprising conversational turns and tangents, whereas remaining grounded within the context like a normal restaurant reserving.
Nonetheless, these early generative fashions have been much less predictable and typically slower than as we speak’s, so we constructed hybrid stacks that paired generative fashions for versatile phrasing with deterministic state monitoring and strict coverage guardrails. This helped make sure that each customers and companies had an excellent expertise.
3. Unbounded duties and Gemini integration
As basis fashions superior, we eliminated the constraint of fastened verticals totally by shifting to prompting the system with free-text process descriptions. This allowed fashions to outline a process merely through a immediate, enabling the system to generalize to new process varieties and nuances (like asking for window seating or additional house for a stroller) and new enterprise classes, with out constructing separate dialogue fashions from scratch.
Our calling stack makes use of Gemini to unify multimodal understanding, planning, and dialogue technology so the system can navigate multi-constraint requests — like scheduling automobile upkeep round calendar conflicts — and adapt when a receptionist suggests one other slot.

