EDITION OF OCTOBER 11, 2026
Voice AI: PolyAI and Otter highlight the limits of automated conversation
Executives at PolyAI and Otter describe the obstacles that persist in voice AI: reasoning speed, transcription, understanding and transparency. Their comments suggest assessing these tools on the reliability of exchanges and actions, beyond the realism of the voice alone.
Legal disclosure: this content was produced by an artificial intelligence system. Editorial transparency
Speaking and listening simultaneously does not resolve all the difficulties facing a voice assistant. Executives at PolyAI and Otter highlight reasoning speed, transcription errors and informing users. Their observations shed light on the conditions for reliable automation, while its actual performance remains to be demonstrated.
Is the quality of a synthetic voice enough to make an assistant useful? Comments from PolyAI and Otter executives reported by TechCrunch shift the debate toward response speed, understanding requests and the reliability of actions. They express assessments by industry participants, without providing comparative measurements of their products.
Shawn Wen, chief technology officer at enterprise voice AI platform PolyAI, believes the sector has yet to reach its tipping point. He acknowledges the development of models capable of speaking while listening, known as full-duplex models. In his view, the next obstacle is to speed up reasoning so that responses arrive quickly enough for the conversation to feel natural.
This distinction matters for customer service applications. Wen believes an agent must inspire enough trust to convince the person it is speaking with that it can solve a problem. His reasoning suggests that acceptance would depend as much on the usefulness of the exchange as on how the voice sounds. This remains a hypothesis about user behavior, however, with no numerical results presented to support it.
At Otter, marketing head Alex Gay emphasizes identifying the people speaking and understanding their intentions, combined with the organization’s knowledge. The company is also working on digital doubles that could represent people in meetings. Gay considers it necessary to preserve emotional expression in the voice to enable exchanges richer than a succession of questions and answers. This work does not demonstrate that the objective has been achieved.
Transcription remains a weak point. Wen believes automatic speech recognition systems sometimes miss important words. Gay stresses that insufficiently accurate transcription can distort the actions that follow and undermine trust. For an organization, this suggests examining the entire chain, from captured speech to the resulting action, rather than assessing only the fluency of a demonstration.
Transparency is another issue raised by the executives. Otter wants to explore notifications in a meeting’s chat to indicate that it is being recorded, including when its bot is not present. Wen also advocates explicitly identifying AI agents in business calls. These intentions outline useful evaluation criteria, but do not establish that they are applied systematically.