Talking to others for you: Leveraging machines that speak the “human conversation” protocol
summary
Demonstrates building production systems that place machines on real phone calls using Google Duplex, shifting from RNNs to end-to-end LLMs and running millions of calls per year for tasks like reservations, business data refreshes and call screening. Uses Pixel features like Hold For Me and Call Screen, measures impact with experiments on a rich conversation corpus, and generates training data via vendors and LLM-driven simulators. Emphasizes data-driven behavior, transparency to the callee, and that the system already serves at scale with trillions of facts seen.
AI-written summary · about
chapters (13)
- 0:00Conversational AI at scale
- 1:13Scaling human-like talks
- 2:37Talking to others for you
- 4:21Challenges in conversational AI
- 6:07Designing natural conversations
- 7:30Goals and call-to-action
- 8:52Learning from real data
- 11:23Treat the other side as user
- 12:50Bootstrapping with low-cost errors
- 14:34Efficient training data
- 16:43Transparency and trust
- 18:30Clarify intentions first
- 19:53Summary and future outlook
related talks
50mSpeak To Me: Voice App Conversation Practices
01h 04mBuilding Expertise Into Generated Conversations
01h 02m[Full Workshop] Building Conversational AI Agents
11mMedia Interview
10mBuilding Conversational AI Agents at Scale
17mBuilding Effective Voice Agents
52mGoing Full Duplex - Building Apps for Speech
32mHarnessing Large Language Models for Enhanced Customer Experience