PolyAI’s CTO Shawn Wen and Otter’s CMO Alex Gay told TechCrunch what a lot of us have been muttering into our headsets for months: voice AI hasn’t had its ChatGPT moment. Not yet. Despite the weekly parade of models claiming to sound human and converse like one, the two of them pointed at speaker identification, intent capture, and emotive expression as the work still left to do.
I test these tools for a living, and hearing that from inside the industry was a small relief. There’s a particular kind of fatigue that comes from downloading a demo, hearing a voice that genuinely does sound like a person for about eleven seconds, and then watching it fall apart the moment the conversation stops behaving like a script.
What “no ChatGPT moment” actually means
ChatGPT’s breakthrough wasn’t that it generated text. Text generators existed. It was that a non-technical person could open a box, type anything, and get something useful back without learning a single trick. The tool met you where you were. No prompt engineering course, no setup, no tolerance for weirdness required.
Voice AI hasn’t crossed that line. Most of what I review still requires the user to adapt. You learn to speak in clean, declarative sentences. You learn not to interrupt. You learn not to say “wait, actually” halfway through a thought because the model will take that as a complete turn and start answering the wrong question. You learn where the tool is brittle and you route around it. That’s the opposite of a mass-adoption moment. That’s a power-user relationship.
Wen and Gay naming speaker identification, intent capture, and emotive expression as the gaps tracks with what breaks in testing. Those three failures look different but they produce the same feeling for a user, which is the sense that nobody is actually listening.
The three gaps, from a tester’s chair
Speaker identification is the one that bites hardest in real meetings. Put four people in a room with overlapping speech, a shared mic, and one person on speakerphone, and transcription quality drops in a way that’s hard to recover from. A transcript that attributes the wrong decision to the wrong person is worse than no transcript, because someone will act on it.
Intent capture is subtler and more consequential. Humans don’t state intent cleanly. We approach it sideways, we revise mid-sentence, we bury the real ask in the fourth clause. A model that only handles the literal surface of an utterance will be technically correct and practically useless. In customer service contexts, where PolyAI operates, that gap is the entire product.
Emotive expression is the one I’d expect people to dismiss as polish, and I think that’s a mistake. Flat delivery isn’t just unpleasant, it strips out information. Emphasis tells you what matters in a sentence. Hesitation tells you something is uncertain. A voice that delivers everything at the same confident register is harder to parse, not just less pleasant to hear.
Why 2026 keeps coming up
Chatter on LinkedIn and in YouTube discussions points to 2026 as the year the meaningful shifts arrive, with natural conversation flow and human-like speech quality as the focus. One of those conversations touched on speech-to-speech as the route forward rather than stitching transcription, text generation, and synthesis together in a chain.
That architectural point is the most interesting thing in this whole discussion, and I’d flag it as something to watch rather than something settled. A pipeline that converts speech to text, processes the text, and converts back to speech throws away a lot on the way in. Tone, pace, hesitation, the emotional texture of how something was said. If the information never enters the system, the system can’t act on it. Going speech-to-speech directly is the obvious fix, but I haven’t tested anything that proves it out at the quality level these executives are describing, so I’m holding my enthusiasm.
What I’d tell you to do about it
Buy for the narrow job, not the demo. Voice AI is genuinely good right now at well-bounded tasks: dictation in a quiet room, single-speaker transcription, routing a caller through a small set of known options. Those tools work and they save real time.
Be skeptical of anything sold on the premise that it converses like a person. That’s the claim Wen and Gay are pushing back on, and they have more commercial incentive to overstate progress than to understate it. When vendors on the inside say the breakthrough hasn’t landed, the marketing copy claiming otherwise deserves a harder look.
The honest read is that voice AI is in its useful-but-finicky phase. That’s not a condemnation. Plenty of tools I rely on daily live in that phase. It just means the thing being promised and the thing shipping are still different products, and knowing which one you’re buying is the whole game.
đź•’ Published: