Indeed, an open-ended interface like a VUI invite unimplemented interactions. I worked on a conversational agent and now realise that there are fundamental usa

Indeed, an open-ended interface like a VUI invite unimplemented interactions. I worked on a conversational agent and now realise that there are fundamental usability barriers that are very difficult to overcome: - Limited information bandwidth. Adults can visually read between 250–600 words per minute, plus they have peripheral vision that can help in scanning for shapes, colors, and images. Plus information in a GUI is persistent on a screen for easy reference. For voice, adults can only comfortably listen at 150–160 WPM and would need to hold information in their memory for reference. This makes voice ecommerce impractical for anything beyond familiar essentials. - Lack of editing layer. It’s simple and straightforward to correct text box input, but it’s difficult to correct a voice command and doing so is ambiguous and adds extra chaotic information. A lot of people think aloud, so they frequently self-correct themselves mid-sentence or add on information in an ad-hoc manner: “I’d like a cappuci—er, make it a latte. (pause) And oh with soy milk.” - High context switching cost. GUI’s keep contexts within frames, tabs, and windows — and the user can easily (low cost) switch between them. This is an Amazon shopping cart context, this is a Reminders context, and so on. A button press is unambiguously contained within its context. In voice, contexts are not quite parallel but sequential and have to be built up over time.