Voice AI Still Lacks Its ChatGPT Moment, Execs Say

Voice has been billed as the next big interface for years, and investors have poured billions into startups building everything from foundation voice models to customer-service agents, meeting notetakers and AI dictation tools. Yet according to two executives who ship enterprise voice products for a living, the technology still has not had its own ChatGPT moment, TechCrunch reported.

What the executives said

Speaking on stage at the HumanX conference, PolyAI CTO Shawn Wen acknowledged that the industry has cleared one real milestone. Full-duplex models, which can listen and speak at the same time, now exist. The next challenge, he said, is making reasoning fast enough that an agent can fetch an answer quickly and keep the conversation feeling natural.

Wen argued that enterprise agents should not sound robotic, and they need to give callers confidence that their problem will actually be solved. Once a caller gets through two or three good turns, he said, they start to trust the agent enough to stop asking for a human.

Alex Gay, CMO of the meeting notetaker Otter, focused on a different gap. Capturing who said what, extracting intent, and then reconciling that with organizational knowledge is the real prerequisite for automation that acts on its own. Otter is exploring digital twins that could represent people in meetings, and Gay stressed that any such output voice has to carry the emotive nuance of a human conversation to be worth anything.

The latency and accuracy wall

The weak link is not the voice itself. Wen said automatic speech recognition models routinely miss important keywords, which corrupts the context the agent reasons over. Gay agreed that transcription is the layer everything else is built on, and that if the first transcript is wrong, every downstream action is flawed too.

Transparency is the other wall. Both executives said tools should tell people when they are being recorded or when they are talking to an AI, rather than letting the interaction slip past unnoticed.

Why it matters for builders

For teams wiring voice agents into production, the takeaway is unglamorous. First, treat reasoning latency as a hard budget: a natural-sounding model that thinks too slowly still feels broken. Second, instrument transcription quality, because ASR errors silently poison every tool call that follows. Third, decide deliberately between cascaded pipelines, which give you control over every stage, and speech-to-speech models, which sound smoother but cost more control. Finally, build disclosure into the flow from day one. Trust, not timbre, is what turns a demo into a system people actually rely on.