Conversation Coach: A Voice-enabled AI System that Helps Practice Difficult Workplace Conversations
| Source: arXiv AI
Tags: voice AI, conversational AI, LLM agents, enterprise AI, EMNLP 2026, speech-to-speech
A voice-first AI coaching system deployed to 40,000+ managers at scale rehearses difficult workplace conversations, with an end-to-end speech model achieving 3x lower latency and 8x lower cost than a cascaded LLM approach, though the cascaded system won out for coaching quality in production.
Details
Conversation Coach is a voice-enabled AI system developed to help managers rehearse difficult workplace conversations — a training need that is costly to deliver at scale via human coaches or text chatbots. The research team at Amazon compared two architectures: a fully end-to-end speech-to-speech model and a cascaded pipeline combining ASR, an LLM, and TTS. The end-to-end approach delivered 3x lower median latency and native barge-in capability at an estimated 8x lower cost. However, the cascaded approach offered superior reasoning quality essential for generating actionable coaching feedback. The team deployed the cascaded architecture in production. Over six months, 40,000+ managers used the system. Adoption patterns suggest selective use — managers tended to engage with it for genuinely challenging conversations rather than routine ones, which the authors interpret as a signal of practical utility rather than curiosity-driven exploration. The paper offers rare production data on voice AI deployment at enterprise scale, including the latency-quality tradeoffs that shaped the architecture choice. Published in the EMNLP 2026 Industry Track.