These executives believe voice AI has not yet reached its ChatGPT moment.

Meta description: PolyAI and Otter outline voice AI progress from full-duplex models to transcription gaps, while investors fund startups building conversational interfaces.

Tags: Voice AI, PolyAI, Otter, HumanX, Customer Service

Featured image: abstract waveform and neural network panels, no real people, no company logos, no text.

Billions of dollars are flowing into voice AI startups as investors bet that spoken interaction will become the next major interface. The category includes model makers, enterprise customer service providers, meeting note-takers, and AI-powered dictation products. New models and tools appear every week claiming human-like sound and conversation, but PolyAI CTO Shawn Wen said at HumanX last month that voice AI has not yet had its ChatGPT moment. Otter CMO Alex Gay highlighted a related challenge: natural output is not enough without accurate context, trust, and disclosure.

From Funding to Full-Duplex Milestones

The funding covers a broad set of products, from the models that generate speech to enterprise services and meeting tools. A usable assistant needs more than a convincing voice. It needs speech recognition, fast reasoning, organizational context, and a way to act on what it hears. Wen said the milestone of developing full-duplex models has been reached; these models can speak while listening to the user. The next challenge, he said, is making reasoning very fast so models can fetch answers quickly and the conversation feels natural.

For enterprise customer service, that timing matters because callers need to trust that the agent can solve problems. Wen said AI agents should not sound robotic. He described a sequence: once the voice is good enough, a customer may engage for the first two or three turns, build confidence, and later feel they do not need to talk to a human if the agent can solve the problem.

Transcription and Context Are the Bottleneck

Gay framed automation around three steps: speaker identification, intent capture, and typing that up with organizational knowledge. Otter is also working on digital twins that might represent people in meetings. For that use, he said the output voice must give the same emotive expressions of talking to a human in a meeting. Without the ability to support debate, strategic discussion, and relationship, an avatar becomes just a question-and-answer chatbot.

The input side is equally fragile. Wen said ASR, or Automatic Speech Recognition, models often miss important keywords, which can prevent the system from capturing the whole context. Gay agreed, adding that Otter keeps working to improve transcription and that language is one area where voice models still need to improve. For Otter, transcription is not the end point; it is the layer that supports productivity gains. If the original transcript lacks the needed accuracy, follow-up actions can become flawed, and users lose trust in the platform.

Disclosure Becomes Part of Trust

New voice tools also create a basic transparency question: do people know they are being recorded or talking to AI? Tools should declare that customers are being recorded or speaking with AI. Otter wants to instill trust among people in a meeting, so even when the bot is not present, it wants 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 in enterprise calls.

The current phase runs from billions of dollars flowing into voice AI to full-duplex models as a visible milestone, then to harder work behind the scenes: faster reasoning, better ASR, and clearer disclosure. What to watch next is whether those parts can combine into conversations natural enough for customers to accept without a human.