Artificial intelligence voice analysis technology can now distinguish between dementia patients and healthy individuals with 90.14% accuracy by evaluating both speech content and non-verbal delivery patterns, according to research presented at Interspeech 2026. Naver Cloud researchers announced the findings on October 7, detailing how tools like OpenAI’s Whisper and GPT-5.2 evaluate speaking speed, pauses, and vocabulary during automated phone calls.
Naver Cloud Demonstrates AI Voice Analysis at Interspeech 2026
Naver Cloud researchers presented two papers at the international speech conference Interspeech 2026 in Australia, showcasing how examining both what people say and how they say it helps identify cognitive health issues. Traditional speech recognition filters out pauses and hesitations as noise. However, the new methodology extracts these non-verbal features to assess health conditions.
The research team utilized OpenAI’s Whisper model to extract speech features. They then analyzed vocabulary, sentence structure, and narrative coherence using the large language model GPT-5.2. Combining both datasets raised the overall performance in distinguishing dementia patients from healthy individuals to 90.14%.
How Pause Duration and Speaking Speed Reveal Cognitive Health
Transcribing speech into plain text obscures critical behavioral markers, such as how long pauses last while recalling words or how frequently interruptions occur. A separate study showed that diagnostic accuracy rose from 81.48% when using only transcribed text to 90.14% once data on pronunciation, pauses, speech speed, and narrative themes were added sequentially.
Naver Cloud previously participated in Local Government and Public Week 2026 held in Tokyo, Japan, in May. During that event, the company showcased ‘Naver Care Call’ and began conducting clinical research to analyze early symptoms of dementia and other conditions using Care Call conversation content.
Clinical Trials Underway at Seoul Boramae Hospital
Naver Cloud is actively conducting clinical trials in collaboration with Emocog and Seoul Boramae Hospital. The ongoing research evaluates whether cognitive decline can be accurately detected through phone conversations alone.
Globally, voice biomarker technology is moving from experimental research into routine medical practice. The Business Research Company, a market research firm, forecasted that the global voice biomarker market would expand from 3.35 billion dollars in 2025 to 7.41 billion dollars by 2030.
Common Questions About AI Voice Biomarkers
What specific speech features does the AI analyze?
The system evaluates vocabulary, sentence structure, narrative coherence, speaking speed, pronunciation data, and the duration and frequency of pauses using OpenAI’s Whisper and GPT-5.2 models.
Who is conducting the clinical trials for this technology?
Naver Cloud is conducting clinical trials alongside Emocog and Seoul Boramae Hospital to test the detection of cognitive decline through phone conversations.
What was the overall diagnostic accuracy of the combined datasets?
Combining transcribed text with non-verbal delivery features such as speech speed and pause duration raised the overall performance in distinguishing dementia patients from healthy individuals to 90.14%.
Related reading