The theory that voice will become the next major computing interface has attracted billions in investment. Startups are building tools for everything from enterprise customer service to meeting note-taking. Every week brings a new model claiming to sound human. However, executives in the field argue that the technology has not yet achieved its defining breakthrough.
Shawn Wen, chief technology officer at enterprise voice AI platform PolyAI, stated that voice AI has not reached its “ChatGPT moment” yet. He made this assessment at the HumanX conference last month. While the industry has achieved full-duplex models that can speak while listening, Wen argues this is only a milestone, not a solution.
Reasoning Speed and Natural Conversation
Wen identified the next major hurdle as the speed of reasoning. He explained that models must fetch answers quickly to make conversations feel natural. If the delay is too long, the interaction breaks down. He also emphasized that AI agents in customer service must not sound robotic. Callers need enough confidence to believe the agent can solve their problems without human intervention.
The goal is to build trust over the first few turns of a conversation. Once a customer feels the agent is competent, they will likely prefer to continue without speaking to a person. This shift requires a level of nuance that current systems often lack.
Transcription Accuracy and Trust
Accuracy remains a critical barrier for tools like Otter, a meeting notetaker. Alex Gay, the company’s chief marketing officer, noted that speaker identification and intent capture are key steps for automation. However, these features rely on a foundation of perfect transcription. If the original transcript is flawed, all downstream actions become unreliable.
Gay pointed out that Automatic Speech Recognition (ASR) models often miss important keywords. This creates gaps in context that prevent the AI from understanding the full picture. When the system takes action based on wrong information, users lose trust in the platform. Otter is working to improve its ASR model to prevent these errors.
Gay also highlighted the importance of emotive expression. He argued that the best meetings involve debate and strategic discussion underpinned by relationships. If an avatar cannot replicate the emotional tone of a human colleague, it remains merely a question-and-answer chatbot. The output voice must convey the same expressions as a human speaker to enable true engagement.
Transparency in AI Interactions
As voice tools become more common, questions of transparency have emerged. Users need to know when they are recording or interacting with an AI. Otter aims to instill trust by notifying meeting participants even when the bot is not actively present. This includes sending chat notifications that the meeting is being recorded.
PolyAI’s Wen agreed that establishing clear boundaries is vital. In enterprise calls, it is important to confirm that people are talking to an AI. This transparency helps manage expectations and ensures that customers are aware of the technology they are engaging with. Without these safeguards, the rapid adoption of voice AI could face significant resistance from users who feel deceived or confused by the technology.
Source: TechCrunch
