The next voice still hasn't had its ChatGPT moment

Hype around voice AI is peaking. Investors are pouring billions into startups building models, enterprise customer-service platforms, meeting-summarization and dictation tools. In practice, though, the technology has yet to deliver the breakthrough that turned ChatGPT into a synonym for generative AI. A new model promising human-like sound and natural conversation is announced almost weekly; the reality on the ground remains far from the promise.

Full duplex is here; reasoning speed isn't

PolyAI chief technology officer Shawn Wen told the HumanX conference last month that the milestone of full-duplex models — systems that can speak while listening — has already been reached. The next hurdle is inference latency: models must retrieve answers fast enough for the conversation to feel natural, without pauses that break the flow. Without that, even a perfect voice remains an empty shell.

Trust first in customer service

AI agents in contact centers don't need to sound robotic, Wen said, but more importantly they must give callers the sense that they can actually solve the problem. After two or three successful exchanges the customer starts to build confidence, and over time may feel no need to speak with a human representative. That shift from curiosity to trust is the real adoption metric, not voice quality alone.

Accurate transcription or nothing

Otter chief marketing officer Alex Gay pointed out that speaker identification, intent capture and cross-referencing with enterprise knowledge form the foundation for genuine automation. The company is also developing "digital twins" to represent people in meetings; for that, the voice must carry the same emotional nuance as human conversation — argument, strategic debate, relationship. Without it, you have a Q&A chatbot in disguise.

Gay emphasized that transcription was never an end goal for Otter; it is the layer on which follow-up actions are built. If the original transcript is inaccurate, every derived action is flawed, and when the platform starts making mistakes, trust collapses. That makes improving the automatic-speech-recognition model critical: the downstream effects are too significant to compromise.

Transparency as a trust prerequisite

The new tools raise the question of transparency: systems must disclose to users that they are being recorded or are speaking with an AI. Otter tries to build trust even in meetings where the bot is not present, for example by posting a chat alert that the meeting is being recorded. Wen added that in enterprise calls it is mandatory to clarify up front that the counterpart is an AI agent; without that, there is no basis for the trust the technology so heavily depends on.