Challenges in Automatic Speech Recognition for Adults with Cognitive Impairment
ASR error rates significantly higher for dementia patients, revealing critical accessibility gap in voice tech.
A new study led by researchers from UC Davis and UC Berkeley reveals a critical accessibility flaw in modern voice assistants. The paper, "Challenges in Automatic Speech Recognition for Adults with Cognitive Impairment," tested Amazon Alexa's ASR system with 83 older adults across three cognitive groups: cognitively normal, mild cognitive impairment, and dementia. The results showed automatic speech recognition error rates were significantly higher for individuals with dementia, creating a substantial usability gap that undermines the promise of voice-enabled smart home systems for supporting daily living with conditions like Alzheimer's.
The research team conducted an acoustic analysis to identify the specific speech features causing ASR failures, finding that a speaker's intensity, voice quality, and pause ratio were key predictors of accuracy. Based on these findings, the authors outline crucial design implications for AgeTech, including developing speaker-personalized ASR models, implementing human-in-the-loop correction systems for transcripts, and creating interaction-level personalization that can adapt to users' changing abilities. This work exposes how current voice AI systems fail vulnerable populations and provides a roadmap for building more inclusive technology.
- Amazon Alexa's ASR showed significantly higher error rates for 83 older adults with dementia compared to cognitively normal peers.
- Acoustic analysis identified intensity, voice quality, and pause ratio as key predictors of ASR accuracy failure.
- Researchers propose three fixes: personalized ASR models, human-in-the-loop transcript correction, and ability-adaptive interaction design.
Why It Matters
Exposes a major accessibility failure in voice AI that affects millions with cognitive impairment, pushing for more inclusive tech design.