type
Post
status
Published
date
Nov 21, 2026
slug
summary
tags
category
技术分享
icon
password
A Decade of Evolution: How Google Duplex Uses Gemini to Make Calls for You
Making phone reservations or booking appointments is often tedious because live conversations are unpredictable. The person on the other end may change their mind mid-sentence, and subtle tone shifts indicate speaking turns. If the AI pauses too long, it seems like the call dropped; if it speaks too soon, it interrupts. To handle this respectfully, the system requires high-quality speech understanding, contextual reasoning, and low-latency speech synthesis working together. We introduced Google Duplex nearly a decade ago to address this, but the architecture has significantly evolved. Recently, we launched Duplex in Tasks in Gemini App, allowing a personal AI collaborator to coordinate real-world chores by calling businesses on behalf of users.
From Narrow Tasks to Shared Infrastructure
What started as a narrow, task-specific system has expanded into a shared conversational infrastructure. During the COVID-19 pandemic, business hours and services shifted rapidly, so Duplex focused on automatically verifying merchant information to keep listings accurate. This remains a critical data source; business information sourced from Duplex calls has been displayed in Search and Maps over 1 trillion times. Simultaneously, we collaborated with Android and hardware teams to bring conversational models directly on-device. Starting in 2019, new Pixel Call Assist features rolled out across five consecutive Pixel generations. These models power features like Call Screen and Hold for Me, which help users filter unknown callers and speak with live agents when hold queues end. More recently, in 2025 and 2026, foundation models enabled broader agentic calling in Google Search and Ask Maps, checking quotes, product availability, and service openings for various businesses.
Three Generations of Conversational Architecture
The stack behind these calls has evolved through three distinct eras, moving from bespoke neural pipelines to a unified architecture using Gemini.
Pre-LLM Closed-Domain Models
Work on Google Duplex began before modern language models and the transformer architecture were invented. To ensure a good experience for both businesses and users, we constrained Duplex to narrow domains and trained custom recurrent neural network models on anonymized, task-specific phone conversations. Without pre-trained world knowledge, the models had to learn how to handle specific task types efficiently. This required extensive training data for each task, such as interpreting the rigid flow of restaurant booking versus checking holiday hours, to ensure reliable navigation of domain-specific logic.
Early Generative Models and Hybrid Stacks
As open-domain foundation models emerged, Duplex began employing them in production. For example, in 2022, Duplex used LaMDA in live calls. These pre-trained models simplified comprehension and common-sense reasoning, allowing the system to handle unexpected conversational turns while remaining grounded in context. However, early generative models were less predictable and sometimes slower than today’s standards. To maintain reliability, we built hybrid stacks that paired generative models for flexible phrasing with deterministic state tracking and strict policy guardrails, ensuring a consistent experience for users and businesses.
Unbounded Tasks and Gemini Integration
With the advancement of foundation models, we removed the constraint of fixed verticals. By shifting to prompting the system with free-text task descriptions, the model can now define tasks simply via a prompt. This allows the system to generalize to new task types, nuances (such as asking for window seating), and new business categories without building separate dialogue models from scratch. Our current calling stack uses Gemini to unify multimodal understanding, planning, and dialogue generation. This unified approach enables the system to navigate multi-constraint requests effectively. By leveraging Gemini, we have moved from closed-domain models to a flexible system that can handle a wide range of real-world interactions, making the process of calling businesses on behalf of users seamless and efficient.
原文出处:Google Research, https://research.google/blog/how-google-duplex-evolved-to-also-call-businesses-for-you-in-gemini/
- 作者:数字Baby
- 链接:myaigc.top/article/41fb7c2d-0e98-8120-943e-ca0a3d6386cd
- 声明:本文采用 CC BY-NC-SA 4.0 许可协议,转载请注明出处。



