Toward Effective Communication of AI-Based Decisions in Assistive Tools: Conveying Confidence or Doubt to People with Visual Impairments at Accelerated Speech

AI-based assistive technologies can mischaracterize information presented to people with visual impairments (PVIs), e.g., by misinterpreting someone's facial expressions. To improve their trustworthiness, algorithmic decisions can be communicated more effectively by conveying their level of confidence to PVIs. Since assistive feedback is typically provided through accelerated audio for PVIs, this work explores how the tone of the audio output can be manipulated to convey confidence or doubt at higher speeds for short sentences. We conducted two online surveys with PVI (n = 151) and sighted (n = 170) participants. We found that PVI participants perceived confidence and doubt for short sentences more accurately at up to 1.5x speedups, whereas, sighted participants perceived them at up to 2x speedups. Other factors such as preferred speedups are also associated with better perception. These results demonstrate the potential for conveying the confidence level of AI-based decisions by choosing the appropriate speedup rates.