Age VoicE, by Fall 2025 Master of Information and Data Science alums Emma Choate, Vinith Kuruppu, and David Russell, aims to serve an aging population by creating a fine-tuned model that will bridge age-related performance gaps in artificial intelligence.
The team was awarded the Fall 2025 Hal R. Varian Capstone Award, which recognizes the semester's top project.
To learn more, we interviewed the team —
What inspired your project?
Emma: Short answer: Amazon Echo SNL skit.
Long Answer: The inspiration for this project came while I was struggling to develop and pitch a capstone idea, and speaking with my mom on the phone. During that conversation, she shared that my grandfather, who lives alone in his two-story, five-bedroom longtime family home, was beginning to show early signs of dementia, including forgetting whether he had let the dog in or out. We as a family worry about his safety, but he values independence, is reluctant to accept help, and assisted living is quite expensive.
“At a broader level, AgeVoicE highlights a simple but powerful idea: accessibility isn’t a niche feature — it’s a standard of engineering quality. When we build systems that work for those most likely to be left behind, we create technology that is more robust, inclusive, and better for everyone.”
This made the wheels spin in my brain. Surely I’m not the only one with an elderly family member (loved one or friend) who was aging at home and could use a little help. I pitched to the class “AI Accessibility for Aging in Place.” The idea was still conceptual, and I had not yet figured out how to turn it into a concrete product. That changed when I met with Dave and Vinith. By joining forces and combining our different professional backgrounds and perspectives, we were able to clearly define both the technical scope and the end goal of the project.
Together, we decided to not only conduct a rigorous evaluation of a state-of-the-art speech recognition system (Whisper) on aging voices, but also show how targeted fine-tuning with aging speech data could meaningfully improve model performance. This collaboration ultimately shaped AgeVoicE into a complete, end-to-end project that combines evaluation, experimentation, and practical impact.
What was the timeline or process like from concept to final project?
Dave: Our capstone began with Emma’s motivation to improve how voice assistants listen to and understand older adults. Her pitch resonated with Vinith and me and shaped our direction for the term. Early weeks were spent on literature review, dataset discovery, and aligning on an approach: fine-tuning the state-of-the-art open-source ASR model, OpenAI Whisper. We were fortunate to have strong data availability, ultimately selecting Mozilla Common Voice for scale and diversity and DementiaBank for its clinically rich representation of older adults and cognitively impaired speakers. With the problem defined, we built our technical stack. We used AWS S3 for storage, EC2 for audio preprocessing, and SageMaker and Google Colab for training, completing data cleaning and loading by week 7.
The second half of the project was defined by iteration, setbacks, and a late breakthrough. Emma led a statistical stress test of the baseline Whisper model, identifying pauses, slowed speech, and especially stutters as key failure modes — patterns common in aging speech. Fine-tuning began in week 8 and proved harder than expected, with hundreds of experiments initially yielding little improvement. In week 12, just before our penultimate presentation, Vinith achieved a meaningful reduction in word error rate.
The breakthrough came when we retrained the full model and paired it with data that better reflected how older adults speak. This allowed the model to relearn how to interpret speech from basic acoustic cues to full word patterns, which was critical for capturing age-related behaviors like pauses, hesitations, and repetitions. By oversampling DementiaBank 5× and preserving clinically significant disfluencies such as filler words and breaks, we trained the model to treat these patterns as signal rather than noise. Once the model began recognizing them as meaningful, accuracy improved substantially — by about 17% — demonstrating that domain-aware fine-tuning and representative data can materially improve voice AI for older adults.
Do you have any future plans for the project?
All: We are actively considering next steps for AgeVoicE. We are exploring the possibility of publishing the research and are eager to see the work continue, because we genuinely care about the objective and believe it can make a real difference.
How could this project make an impact, or, who will it serve?
Vinith: We designed AgeVoicE to bridge a critical equity gap in modern AI systems. While state-of-the-art speech recognition models are transformative for many, they consistently underperform on atypical speech patterns, effectively silencing the populations who could benefit from them most.
We see AgeVoicE serving 3 main groups:
Aging adults and individuals with speech differences: For aging adults and those navigating cognitive decline, voice technology is more than a convenience; it is a tool for autonomy. AgeVoicE ensures these individuals can interact with digital ecosystems with confidence, dignity, and independence.
Families, caregivers, and clinicians: AgeVoicE reduces friction in communication, documentation, and care coordination — especially in clinical settings where accurate speech recognition can save time and prevent misunderstandings.
Research and development community: AgeVoicE provides developers building assistive technologies with a practical, scalable blueprint for making modern speech systems more robust to real-world diversity.
At a broader level, AgeVoicE highlights a simple but powerful idea: accessibility isn’t a niche feature — it’s a standard of engineering quality. When we build systems that work for those most likely to be left behind, we create technology that is more robust, inclusive, and better for everyone.
