Analysis
Voice AI hasn't had its "ChatGPT moment," according to executives from two companies building it. Speaking separately -- PolyAI CTO Shawn Wen at the HumanX conference in September, and Otter CMO Alex Gay in comments to TechCrunch -- both argued the category still falls short of the trust and fluency threshold that made ChatGPT a mainstream product almost overnight.
Full-Duplex Is Solved. Trust Isn't.
Wen says full-duplex models -- ones that can speak while listening, instead of waiting for a pause -- are a milestone the industry has already reached. The harder problem now is making reasoning fast enough that responses arrive without an awkward lag, and making a customer-service agent's tone sound non-robotic enough that callers extend it trust over several turns. Wen's bet is that once a caller trusts an agent through the first few exchanges, they stop missing a human -- but getting to that first few exchanges reliably is still unsolved.
“Wen says full-duplex models -- ones that can speak while listening, instead of waiting for a pause -- are a milestone the industry has already reached.”
Gay's version of the same problem sits upstream of tone: automatic speech recognition still misses important keywords often enough to break context, and every transcription error compounds through whatever the agent does next. Otter is building toward "digital twins" that could represent people in meetings with emotive, human-like voice output, since Gay argues meetings built on debate and relationships don't work with a bot that only answers questions -- but that product depends on first getting transcription and speaker identification right, which the company still treats as its core technical risk rather than a solved input.
Both executives raised transparency as a non-negotiable rather than a nice-to-have: enterprise callers should know when they're talking to AI, and Otter wants to notify every meeting participant when AI is recording, even when its bot isn't physically present. That's a compliance and product-design constraint that competes directly against the goal of sounding indistinguishable from a human.
The framing matters for anyone pricing voice-AI deals right now. The category has pulled in, per TechCrunch's framing, "billions of dollars" of venture money on the promise of a ChatGPT-style adoption curve, but the two executives closest to shipping product are the ones saying that curve hasn't started. ElevenLabs, Vapi, Bland and a long list of call-center AI startups are all racing toward the same trust threshold Wen describes -- and until word-error rates and latency both clear it reliably under real call-center noise, rather than demo conditions, the gap between "impressive demo" and "I don't need a human" stays open.
