Voice AI Still Short of ChatGPT‑Level Adoption, Execs Warn

PolyAI CTO Shawn Wen cautioned that voice AI has not yet achieved the ChatGPT moment, saying the industry has reached full‑duplex models but still struggles with rapid reasoning.

Key points

• PolyAI’s full‑duplex models can speak while listening, yet reasoning speed remains a bottleneck.
• Otter emphasizes speaker ID, intent capture, and transcription accuracy for productivity.
• ASR errors often miss key terms, breaking downstream context.
• Transparency protocols must disclose AI presence and recording status.
• Emotional nuance is lacking, limiting engagement in meetings and customer service.

What happened

At the HumanX conference last month, Wen announced that PolyAI had achieved full‑duplex speech but noted, "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". He added that agents should sound natural enough for customers to skip human touch after a few turns.

Concurrently, Otter’s CMO Alex Gay highlighted the company’s focus on speaker identification, intent capture, and accurate transcription. Gay said, "If you think about the meetings that you’re in right now, conversations that you have are where you can have debate, and strategic discussions," stressing that avatars lacking emotional depth feel like a Q&A chatbot.

Both executives point to the same bottleneck: Automatic Speech Recognition models miss critical keywords, eroding context for downstream actions and eroding trust.

The backstory

The voice‑AI sector has attracted billions of dollars as investors chase the next big interface. Every week a new model promises human‑like conversation, yet enterprise deployments often falter on context, accuracy, and transparency. Full‑duplex technology arrived last year, but reasoning latency and ASR fidelity remain the primary hurdles.

Why voice AI lags

Voice models excel at generating speech but lag in real‑time inference. The need to fetch answers quickly forces a trade‑off between model size and latency. Smaller, faster models sacrifice reasoning depth, while larger models exceed acceptable response times for natural dialogue. This mismatch explains why many enterprise assistants still sound robotic and struggle to maintain context.

Implications for enterprise trust

If transcription accuracy is low, every follow‑up action—calendar invites, task assignments, or support tickets—becomes unreliable. Users lose confidence in the platform, and the cost of re‑work rises. Moreover, without clear disclosure that a conversation involves AI, organizations risk violating privacy regulations and eroding employee trust.

What builders and businesses should do

• Prioritize ASR accuracy: invest in continuous model fine‑tuning and speaker‑specific data.
• Measure inference latency as a core KPI; aim for sub‑second response times in high‑stakes contexts.
• Embed transparency flags: automatically notify participants when an AI is active or recording.
• Design emotional cues into avatars: incorporate prosody and pause patterns that mirror human conversation.
• Build feedback loops: capture user corrections to improve intent recognition and transcription.

The bottom line

Voice AI is still chasing the naturalness of ChatGPT. Expect the next wave of models to focus on inference speed and transparency, with enterprise adoption accelerating by 2027 as trust gaps close.