Voice AI is attracting investment and new tools promise conversations that sound increasingly human. But executives at enterprise voice platform PolyAI and meeting notetaker Otter say the technology still has work to do before people can rely on it in calls and meetings.
Natural speech needs fast answers
PolyAI CTO Shawn Wen says full-duplex models, which can speak while listening, mark progress but have not yet given voice AI its “ChatGPT moment.” The next hurdle is making models reason quickly enough to retrieve answers without disrupting the flow of a conversation.
That speed matters because a voice exchange depends on more than a convincing-sounding voice. If a system takes too long to respond, the interaction may feel unnatural even when the speech itself sounds human. For customer service, the agent also needs to help callers feel confident that it can resolve their problem.
Wen described confidence as something that may grow over the first two or three turns. If customers find that an AI agent can solve their issue, they may become more willing to continue without asking for a human. That makes the quality of the exchange part of the service itself.
Meeting assistants need to understand context
For Otter CMO Alex Gay, useful meeting automation begins with identifying who is speaking, capturing what they intend to say and connecting that information with organizational knowledge. Otter is also working on digital twins that could represent people in meetings.
Gay said those representations need more than spoken answers. Meetings often include debate, strategic discussion and relationships between participants. If an avatar cannot support those qualities, he argued, it risks becoming little more than a question-and-answer chatbot.
This points to a broader challenge for voice tools: sounding fluent is not the same as taking part in a meaningful conversation. Systems used in meetings need to preserve enough of the discussion’s substance and human expression to be useful to the people involved.
Transcript mistakes can undermine trust
Voice assistants and meeting tools can misunderstand users, produce incorrect transcripts or generate summaries that miss the mark. Wen said automatic speech recognition models may overlook important keywords, leaving out information needed to understand the full context.
Gay said transcription is a foundation for Otter’s productivity features, rather than an end in itself. When the original transcript is inaccurate, later actions built on it can also be flawed. If a system then acts on that flawed information, users may lose trust in the platform.
The consequences can travel through a whole workflow. A missed word can affect a summary, and a faulty summary can affect what a tool does next. That is why the executives see continued improvement to speech recognition and language understanding as central to making voice AI dependable.
People should know when AI is listening
Accuracy is only part of the trust question. People also need to know when a meeting is being recorded or when they are speaking with an AI. PolyAI’s Wen said callers should be made aware that an enterprise call involves AI.
Otter is considering ways to notify everyone in a chat that a meeting is being recorded, including meetings where its bot is not present. The aim, Gay said, is to build trust among meeting participants.
Together, these concerns set a practical bar for voice AI: respond quickly, understand what people said, support the context of the conversation and be transparent about recording and AI use. The executives’ comments suggest that the technology’s next advance depends on making all of those parts work together.