Voice AI Technology Faces Hurdles Before Gaining Trust

Industry insiders argue that voice artificial intelligence (AI) technology still has a long way to go before it can be trusted in real business environments. Despite rapid advancements, two key barriers must be overcome: inference speed and speech recognition accuracy.

The ability of AI to converse naturally with humans is not enough; it also needs to accurately understand user requests and solve problems. Executives from global voice AI companies emphasized that the technology has yet to experience its ‘ChatGPT moment’, which drove generative AI into the mainstream.

Shawn Wen, Chief Technology Officer at PolyAI, an enterprise voice AI platform, pointed out that full-duplex models have been developed, allowing for natural conversations. However, the next challenge is making inference extremely fast so that answers can be retrieved quickly and conversations feel natural.

The development of full-duplex models has enabled AI to listen and respond while a speaker is still talking. This marks a significant improvement over previous systems where users had to finish their questions before AI would begin answering.

However, the ability to exchange words naturally and answer questions quickly are separate issues. Even if AI responds instantly, finding information for an answer takes time, making it difficult to maintain natural dialogue flow when pausing mid-conversation to search for answers.

Wen used customer service centers as an example, noting that users will immediately seek human agents if an AI agent fails to demonstrate problem-solving ability within the first two or three exchanges. Building trust in early interactions may eliminate the need for users to request human assistance.

Speech recognition errors can undermine all downstream work and lead to misinterpreting user requests. Automatic speech recognition (ASR) models frequently miss critical keywords, causing problems in grasping overall context. This highlights the importance of accurate understanding of human speech for voice AI to handle real business tasks.

Alex Gay, Chief Marketing Officer at Otter, emphasized that if meeting content is transcribed incorrectly, errors can propagate not only to summaries but also to work conducted based on those summaries. He stressed that ‘the moment AI takes a wrong action, users lose trust in the platform’ due to inaccurate transcription.

Otter views transcription as a foundational layer for driving productivity gains, but if it’s inaccurate, every automation feature built on top of it becomes unstable. The company is developing digital twin technology that allows AI to attend meetings and engage in discussions with participants. However, this requires accurately capturing conversational context and grasping relationships between participants.

Gay noted that if AI avatars cannot implement discussion and interaction based on these relationships, they are no different from simple question-answering chatbots. The goal is to move beyond recording remarks to understanding conversational flow and engaging in debate. This would enable digital twins to replace human roles in actual meetings.

The industry recognizes the importance of combining speaker identification, intent capture, and organizational knowledge to generate records that can be used for automation. Transparency concerns are also emerging as new voice tools proliferate; users must be notified when recording is in progress or conversing with AI.

Otter is testing an approach where participants are notified via chat even in meetings without a bot present that the meeting is being recorded. PolyAI’s Wen emphasized the need to clearly disclose to the other party during enterprise calls that they are speaking with AI. This highlights the growing demand for transparency and trustworthiness in voice AI technology.

Despite intense investment enthusiasm, there is a perception within the industry that actual performance is not keeping pace with marketing claims. Billions of dollars are flowing into startups across the spectrum from model developers to enterprise customer service providers, meeting transcription tools, and AI dictation services.