Best Miami News connects businesses and publishers

collapse
Home / Daily News Analysis / These execs think voice AI hasn’t reached its ChatGPT moment yet

These execs think voice AI hasn’t reached its ChatGPT moment yet

Oct 29, 2026  Twila Rosenbaum 38 views
These execs think voice AI hasn’t reached its ChatGPT moment yet

Voice has been billed as the next major computing interface for years, and the recent wave of investment has made that prediction feel more concrete. Investors have poured billions of dollars into voice AI startups, spanning foundation model developers, enterprise customer service platforms, meeting note-takers, and AI-powered dictation tools. Every week brings another model or product release promising speech that sounds human and conversations that flow naturally.

But the gap between demo and deployment remains wide. Two executives working on commercial voice AI said during a recent industry conference panel that the sector has not yet had its defining breakthrough. They described impressive technical progress, especially in full-duplex models that can speak while listening, but argued that the harder problems are still ahead. Those problems include fast reasoning, accurate transcription, emotional nuance, and the kind of transparency that lets customers know whether they are talking to a person or a machine.

Key Facts

  • PolyAI chief technology officer Shawn Wen says full-duplex models represent progress but not a ChatGPT moment for voice AI.
  • Wen argues the next technical challenge is making reasoning fast enough that answers arrive quickly and conversations feel natural.
  • Otter chief marketing officer Alex Gay says speaker identification, intent capture, and connecting transcripts to organizational knowledge are essential for automation.
  • Otter is exploring digital twins that can represent people in meetings, but Gay says such avatars must convey human-like emotion and support debate, not just Q and A.
  • Both executives say automatic speech recognition errors remain a major problem, because inaccurate transcripts undermine downstream actions and trust.
  • The companies also emphasize transparency, including telling people when they are recorded or speaking with an AI agent.

Full-Duplex Is Progress, Not the Finish Line

Full-duplex voice models are a significant technical advance. Unlike turn-based systems that wait for a user to finish speaking before processing a response, full-duplex systems can listen while they speak. That allows them to handle interruptions, backchannel cues, and the kind of overlap that happens in ordinary human conversation. The result can feel less stilted and more responsive.

But PolyAI CTO Shawn Wen said during a recent industry conference panel that full-duplex capability is a milestone, not a destination. The harder challenge is making reasoning fast enough. A voice agent must not only hear the words and generate a human-sounding reply; it must understand the request, retrieve the right information, apply business rules, and decide what to do next, all within a few hundred milliseconds. If the system pauses too long, the conversation feels unnatural. If it answers too quickly with the wrong information, the customer loses confidence.

That balancing act is especially difficult in enterprise settings. A customer service agent may need to check an order status, review a refund policy, verify an account, and update a CRM record while the caller is waiting. Each step adds latency. The model must also know when to ask a clarifying question, when to escalate, and when to stop talking. Full-duplex listening helps with natural turn-taking, but it cannot compensate for slow reasoning or poor data access.

A ChatGPT Moment Would Require More Than Natural Speech

The phrase ChatGPT moment has become shorthand for a technology crossing from novelty to mass adoption. For voice AI, that moment would require more than a convincing demo. It would require systems that people can rely on for everyday tasks without constant frustration. The technology would need to understand accents, background noise, interruptions, and incomplete sentences. It would need to remember context across a conversation and act on that context correctly.

Users also expect voice systems to respond quickly. A conversational agent that takes several seconds to answer feels broken, even if the underlying model is sophisticated. In customer service, long pauses lead callers to hang up or demand a human. In meetings, slow transcription and summarization reduce the value of the tool. The technical stack has to bring together speech recognition, language understanding, reasoning, and speech synthesis within a tight latency budget.

That means voice AI is not one model but a chain of components. A weakness in any link can undermine the whole experience. A strong language model cannot fix a missed account number. A natural-sounding voice cannot rescue a wrong refund promise. A fast response cannot build trust if the system fails to disclose that it is automated. The breakthrough, when it comes, will be judged by reliability rather than by how human the voice sounds in a short clip.

Customer Service Tests Trust Turn by Turn

Wen said AI agents in customer service should not sound robotic and should give callers enough confidence that their problems can be solved. The first two or three turns are critical. If the voice is good enough and the agent handles those early exchanges well, the customer starts to engage. Over time, they may feel they do not need to talk to a human if the agent can solve the problem.

That confidence is fragile. One misunderstood account number, one incorrect policy answer, or one failed handoff can destroy it. Enterprises therefore need guardrails, escalation paths, and audit trails. Voice AI can reduce wait times and handle routine requests, but only if it recognizes its limits. The goal is not to imitate a human perfectly. The goal is to solve problems reliably and make the interaction easy to complete.

The commercial case is strong. Customer service is expensive, repetitive, and often overloaded during peak periods. Voice agents can handle password resets, order status checks, appointment changes, and basic troubleshooting around the clock. But those same tasks are where errors are most visible. A successful deployment requires tight integration with back-end systems, clear escalation rules, and continuous testing against real call recordings. Without those foundations, the agent may sound impressive while failing at the job.

Meeting Agents Need Context, Not Just Transcripts

Otter CMO Alex Gay said speaker identification, intent capture, and the ability to connect a transcript with organizational knowledge are key steps toward automation. A meeting note-taker that only produces a wall of text is useful, but limited. The more valuable system identifies who said what, captures decisions and action items, and links them to the right project, customer, or internal document.

The company is also working on digital twins that might represent people in meetings. For that technology, Gay said it is paramount that the output voice conveys the same emotive expressions as talking to a human. The best meetings involve debate, strategic discussion, and a sense of relationship. If an avatar cannot participate in that kind of exchange, it is just a Q and A chatbot.

That distinction matters because meeting agents are moving from transcription into participation. A digital twin could attend a routine update, summarize a discussion, or answer questions on behalf of a colleague. But representing a person raises questions about consent, identity, and accuracy. The technology must make clear when a twin is speaking and what authority it has. It also must avoid inventing positions that the person would not take.

Transcription Errors Cascade

Wen said ASR models often miss important keywords, which creates a problem for capturing the whole context. In a customer service call, a missed word can change the meaning of a request. In a meeting, a missed name or number can turn an action item into a mistake. Gay agreed that transcription accuracy remains critical, adding that language is an area where voice models need to improve.

For Otter, he said, transcription was never the end point. It was the layer that could drive productivity gains. But if the original transcription lacks accuracy, every follow-up action becomes flawed. Once an automated system takes a wrong action, users lose trust in the platform. That is why improving the ASR model is not a back-office concern; it has significant downstream impacts.

In practice, voice systems must handle accents, jargon, background noise, multiple speakers, and multilingual conversations. They need speaker diarization, custom vocabularies, and confidence scores that flag uncertain passages. In high-stakes workflows, the system should route low-confidence transcripts to a human reviewer rather than silently guessing. The same logic applies to voice agents: they need to know when they do not know.

Transparency Is Becoming a Product Requirement

With new voice tools, there is also a question of transparency. Tools should declare to customers that they are being recorded or talking to AI. Otter said it wants to instill trust in people who are in a meeting, so even for a meeting where the bot is not present, it wants to try methods like notifying everyone in the chat that the meeting is being recorded. PolyAI’s Wen also said it is important to establish that people are talking to an AI in enterprise calls.

Transparency has both legal and product dimensions. Several jurisdictions require notice before recording calls, and consumers are increasingly sensitive to being deceived by automated systems. A voice agent that pretends to be human may work in a short demo, but it can damage trust once the customer discovers the truth. Clear disclosure at the start of a call, visible recording notices in meetings, and simple opt-out paths can reduce that risk.

For enterprise buyers, transparency is also a governance requirement. They need logs, audit trails, and controls that show what the AI said and why. They need to know whether a digital twin attended a meeting and what it contributed. They need to be able to turn the system off. The next phase of voice AI will not be won by the most convincing imitation of a human voice alone. It will be won by systems that solve problems reliably, disclose their role clearly, and know when to hand the conversation back to a person.


Source:TechCrunch News


Share:

Your experience on this site will be improved by allowing cookies Cookie Policy