📰 Key Takeaways

Voice AI is pulling in massive investment, but industry executives say the category hasn’t had its ChatGPT moment yet. Shawn Wen, CTO of enterprise voice AI platform PolyAI, said at the HumanX conference that even though full-duplex models (which can listen and speak at the same time) now exist, voice AI still hasn’t hit a real breakthrough point. He thinks the next key challenge is making model inference dramatically faster, so systems can retrieve answers quickly and conversations feel natural. He also stressed that AI agents in customer service shouldn’t sound robotic — callers need enough confidence that their problem can actually get solved. Once the voice quality is good enough and a customer is willing to keep engaging through two or three exchanges, trust builds gradually, and eventually they feel like they don’t need to be transferred to a human agent.

Alex Gay, CMO of meeting-notes tool Otter, argued that speaker identification, intent capture, and tying conversations back to organizational knowledge are the key steps toward automation. The company is also building “digital twins” to represent real attendees in meetings, and it sees preserving human-like emotional expression in the output voice as critical to making that work. Gay pointed out that the meetings that actually matter are usually built on debate, strategic discussion, and interpersonal dynamics — if a digital twin can’t do that, it’s just a Q&A bot.

Both execs also flagged a corrosive problem for voice AI around understanding and transparency: automatic speech recognition (ASR) models often miss key words, leaving gaps in context. Otter admitted that transcription accuracy is the foundation everything else — summaries, action items — gets built on. Once a transcript goes wrong, the downstream automation goes wrong with it, which erodes users’ trust in the platform.

💬 JudyAI Lab’s Take

Investment in voice AI keeps heating up, but executives are openly admitting the “ChatGPT moment” hasn’t arrived yet — and that gap is worth watching closely.

PolyAI’s CTO points out that full-duplex models already exist, but inference speed is still the bottleneck — systems need to retrieve answers fast for conversation to feel natural, and AI agents in customer service can’t sound robotic. Customers need two or three good exchanges before they build enough trust to stop wanting a human. Otter’s CMO adds that speaker identification, intent capture, and tying conversations to organizational knowledge are the real steps toward automation — and for its “digital twins” to work, the output voice has to preserve emotional expression, or the avatar can’t carry meetings that depend on debate and strategy. Both flag that ASR dropping key words erodes context and user trust — which tells you voice AI’s core bottleneck isn’t how human the generation sounds, it’s how precise the understanding is.

If you’re evaluating a voice AI product, look at transcription accuracy first — that’s the foundational metric — not just how natural the voice sounds.

📅 Source Info

🔗 Related Reading