Naver Cloud Releases Two AI Studies for Detecting Early Signs of Dementia via Voice

Naver Cloud has released two AI research achievements that detect early signs of dementia through voice analysis. These studies were accepted to the international speech conference Interspeech 2026, achieving identical results that exceeded previous best performance through different approaches.

Two Approaches

The first study, "Listening Between the Lines," adopted an approach combining acoustic information extracted from the internal encoder of the Whisper speech recognition model with linguistic information analyzed by an LLM. After separating the acoustic and linguistic pathways, the two signals were combined using a gated fusion network to improve classification accuracy.

The second study, "LoRA-Tuned Large Language Models for Dementia Detection via Multi-View Speech-Derived Features," takes an approach that entrusts all judgment to a single LLM without complex combination structures. Information from four perspectives—transcribed text, pause markers, speech flow statistics, and phoneme sequences—was composed into a single prompt and input into Qwen3 and Gemma3 series models, trained using the LoRA technique.

Performance and Implications

Both studies recorded an F1 score of 90.14% on the public dataset ADReSSo benchmark, surpassing the previous best record of 87.32%. Study ① demonstrated that the fusion of acoustic and linguistic information improved performance by 7.1%p and 14.1%p respectively, while Study ② confirmed that discourse-level topic information contributed the most (5.81%p) to performance improvement.

These results suggest that the key lies in reading both 'what is said' and 'how it is said' together, rather than relying on specific model structures. Naver Cloud aims to build a safety net that captures signs of cognitive health abnormalities through a single phone call and connects them to early medical care by integrating this technology with the CLOVA CareCall service.