Amit Desai distinguishes recognition accuracy from the decision a voice assistant makes under uncertainty. Using a simplified music-request example, he compares acting immediately with stopping to ask the user to repeat. A calibrated confidence distribution provides the basis for choosing thresholds rather than guessing them.
Amit Desai introduces an outcome user cost heuristic that assigns different costs to an incorrect action, a repeated request and a confirmation. He adds a confirm option between act and stop, then illustrates how minimizing the weighted cost can improve the experience even when the underlying accuracy stays fixed. The numerical thresholds are example-specific, not universal settings.
Amit Desai extends the principle to multimodal television interfaces and embodied assistants. The cost of a mistaken action depends on the device, available correction methods and consequences. He argues that real systems need learned, context-sensitive decision behavior alongside improvements to recognition models.
Watch on YouTube
