The race to make voice the next major computing interface is attracting billions of dollars in investment, with startups building foundation models, enterprise customer-service agents, meeting note-takers and AI-powered dictation tools. New products arrive almost every week, many promising conversations that sound human and flow naturally. Yet the executives building these systems say the technology has not yet delivered its defining breakthrough.
PolyAI CTO Shawn Wen argues that voice AI has not reached its “ChatGPT moment,” even though the industry has made real progress on full-duplex models that can listen and speak at the same time. Otter CMO Alex Gay says meeting tools still need sharper speaker identification, intent capture and emotional expression before they can support deeper automation. Their comments point to a common theme: voice AI has produced impressive demonstrations, but reliability and trust remain unfinished work.
Key Facts
- Voice AI has not yet had its “ChatGPT moment,” according to PolyAI CTO Shawn Wen.
- Full-duplex voice models can speak while listening, but fast reasoning is still needed to make conversations feel natural.
- Otter is working on speaker identification, intent capture, organizational knowledge and digital twins for meetings.
- Automatic speech recognition errors can undermine transcription accuracy and destroy trust in downstream automation.
- Companies say they must disclose when calls are recorded or when users are talking to AI.
Full-Duplex Is Only One Milestone
For years, voice assistants struggled with the basic rhythm of conversation. Users had to wait for a prompt, speak in a narrow window and tolerate awkward interruptions. Full-duplex models changed that dynamic by allowing a system to listen while it speaks, making exchanges feel more like a phone call and less like a walkie-talkie. Wen said the industry has reached that milestone, but he does not view it as the finish line.
“We have reached the milestone of developing full-duplex models. The next challenge is to make reasoning very fast, so that the models can fetch answers quickly and the conversation feels natural,” he said during a recent industry conference appearance.
That emphasis on speed reflects how people judge voice systems. A pause that feels brief in text can feel uncomfortable on a call. If a customer asks about a billing issue, a refund or a technical problem, the agent must retrieve information, reason through policy and respond without long silences. The model must also handle interruptions, corrections and changes in intent. Wen said customer-service agents should not sound robotic, and they must give callers enough confidence that their problems can be solved.
He described trust as something that builds over the first few turns. Once the voice quality is good enough and the customer is willing to engage for two or three exchanges, confidence begins to grow. Over time, customers may feel they do not need a human agent if the AI can resolve the issue. That shift depends less on a single model release and more on consistent performance across thousands of real conversations.
Otter’s Ambitions Go Beyond Transcription
Otter’s Gay sees speaker identification, intent capture and the ability to combine those insights with organizational knowledge as key steps toward automation. A meeting transcript is useful, but a system that knows who said what, what was decided and which follow-up actions matter can become part of the workflow. Otter is also exploring digital twins that might represent people in meetings.
For that technology, Gay said the output voice must convey the same emotional expressions as a human conversation. “If you think about the meetings that you’re in right now, the best conversations that you have are where you can have debate, and strategic discussions, and when you feel like there’s a relationship that underpins it. If you aren’t able to have that with an avatar, then it’s just a Q and A chatbot,” he noted.
The comment highlights a gap between information retrieval and genuine collaboration. A meeting assistant can summarize a discussion, but strategic work often depends on nuance, pushback, humor and relationship-building. A digital twin that cannot reproduce those qualities may struggle to earn a seat at the table. Voice becomes the interface, but emotional intelligence and contextual awareness become the product.
Accuracy Is a Trust Problem
Voice AI models have improved rapidly, but users still encounter assistants that misunderstand requests, meeting note-takers that produce wrong transcripts and summaries that miss the point. Wen said automatic speech recognition models often miss important keywords, which creates problems for capturing the full context of a conversation.
Gay agreed, saying Otter continues to work on improving transcription and that language remains an area where voice models need to improve. “For Otter, you know, transcription was never the end point. It was just the layer that we could start to drive some of the productivity gains on the back of. But if your original transcription didn’t have the accuracy that you needed, all the follow-up actions that you have become flawed. And the minute that starts to take action, that is wrong. You lose trust in the platform. It is critical for us to continue to improve that ASR model because all of the downstream impacts are significant,” he said.
That trust problem is especially costly in enterprise settings. A wrong transcript can lead to an incorrect task assignment. A misunderstood customer request can trigger the wrong refund, cancellation or escalation. A meeting summary that misstates a decision can spread through an organization. Voice AI vendors therefore have to treat accuracy not as a technical metric but as a product requirement that determines whether users keep relying on the system.
Transparency Becomes a Product Requirement
The new generation of voice tools also raises questions about disclosure. Tools should tell customers when they are being recorded or when they are talking to AI. Otter says it wants to instill trust in people who are in a meeting, so even when the bot is not present, it wants to use methods such as notifying everyone in the chat that the meeting is being recorded. Wen said it is important to establish that people are talking to an AI during enterprise calls.
Transparency is not only an ethical issue. It affects adoption. Customers who feel tricked may reject a voice agent even if it resolves their problem quickly. Employees who are unsure whether a meeting is being recorded may change how they speak or resist the tool altogether. Clear disclosure gives users a sense of control and reduces the risk that a bad experience turns into a broader backlash.
Why the ChatGPT Moment Has Not Arrived
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 work across accents, background noise, poor connections and emotional conversations. It would require low latency, accurate retrieval and the ability to recover gracefully when something goes wrong.
Voice is inherently harder than text. A text interface can wait while a model thinks. A voice interface cannot. Users hear every pause, every filler word and every sign that the system is unsure. They also expect turn-taking to feel natural, which means the model must know when to speak, when to listen and when to interrupt. Full-duplex models help with overlapping speech, but they do not solve the reasoning bottleneck. The system still needs to fetch answers quickly and decide what to say next.
Enterprise deployment adds another layer of complexity. Customer-service agents must connect to billing systems, order databases and knowledge bases. They must know when to escalate to a human. Meeting assistants must integrate with calendars, chat platforms and project-management tools. Digital twins must protect privacy and avoid misrepresentation. Each integration creates another failure point, and each failure can erode trust.
Investors continue to pour money into voice AI because the interface promise is enormous. A reliable voice agent could reduce call-center wait times, help workers document meetings and make software more accessible. But the executives closest to the technology are signaling that the market is still in a build-and-prove phase. The winners will not be determined by who has the most human-sounding voice clip. They will be determined by who can deliver consistent outcomes at scale.
What Would a Real Breakthrough Look Like?
The executives’ comments suggest several markers for a true voice AI breakthrough. Answers would arrive quickly enough to feel conversational. Speech recognition would capture names, numbers, product terms and intent without constant correction. Meeting tools would identify speakers accurately and turn conversations into reliable actions. Digital avatars would convey emotion without pretending to be human. And every system would make clear when it is recording, when it is generating and when a person should take over.
For now, the gap between a polished demo and a trusted daily tool remains the central challenge. Voice AI has moved beyond rigid menus and robotic prompts, but it has not yet become the invisible, dependable interface that its most ambitious backers describe. The next phase will be judged by latency, accuracy, transparency and the willingness of users to keep talking.
Source:TechCrunch News