Utilizing Artificial Intelligence (AI) Techniques to Decode Healthy Speech Patterns in South America: Study Design and Preliminary Results

María Carello, Fleni, CABA, Argentina
Greta Keller, Fleni, CABA, Argentina
Lara Gauder, CONICET, CABA, Argentina
Laouen Belloli, CONICET, CABA, Argentina
Gustavo Juantorena, CONICET, CABA, Argentina
Nicolás Corvalán, Fleni, CABA, Argentina
María Bellucci, Fleni, CABA, Argentina
Iris Worff, Fleni, CABA, Argentina
Ricardo Allegri, Fleni, CABA, Argentina
Lucía Crivelli, Fleni, CABA, Argentina
Diego Fernandez Slezak, CONICET, CABA, Argentina


Early detection of Mild Cognitive Impairment (MCI) is of paramount importance for the development of straightforward preventive interventions. Thus, there is a significant need for concise and cost-effective cognitive screening tools to identify individuals who warrant further evaluation. Traditionally, language and speech analysis has relied on expert opinion, clinical assessment, and manual linguistic analysis performed by professionals. However, recent advances in Artificial Intelligence (AI) enable instant and accurate extraction of linguistic, facial, and acoustic features. The predictive potential of using these features is still in its infancy in MCI research, and the application of this tool remains unexplored in Argentina. The main objective of this study is to develop and validate a neuropsychological tool based on natural language processing techniques. In this context, we aim to present preliminary data on language patterns observed in healthy individuals.

Participants and Methods:

We recruited 29 healthy individuals from Fleni Memory and Aging Clinic in Buenos Aires, Argentina. We conducted a comprehensive assessment that included neuropsychological tests (Uniform Data Set - 3) and magnetic resonance imaging (MRI) following the ADNI 3 protocol. Participants were recorded in image and voice while performing a language task. They were required to describe two target images (in a randomized order), the 'Cookie Theft' and 'Firefighter-Oasis.'

Additionally, as a baseline measure, the description of their favorite sandwich and reading of the cheese story was also required. We employed automated speech transcription to analyze data and conducted an in-depth analysis of acoustic features, utilizing metrics to assess coherence, fluency, sentiment, and speech graphs. Furthermore, we utilized video data to evaluate facial muscle movements using the face mesh technique (Lugaresi et al., 2019).


We analyzed silence length and speech across tasks. The description tasks showed significant differences in their length (t = -2.70, p < 0.01) but not in the silences (t = -1.82, p = 0.07). During the description of a "Firefighter," the average silence length was 29.9 seconds (±12.10), while during the description of the "Cookie Theft," it was 22.40 seconds (±10.14). When describing their favorite sandwich, the average silence length was 17.82 seconds (±9.66), and during the reading of the cheese story, it was 20.35 seconds (±6.75).


This preliminary study aims to enhance our understanding of the role of speech patterns in performance tasks. These patterns may exhibit greater prominence within clinical populations, facilitating a more effective differentiation between the two groups. It is crucial to note that these results are preliminary and are based on healthy individuals. In future phases of the research, we plan to include patients with MCI to gain a comprehensive understanding of communication patterns within this clinical context.

Category: MCI (Mild Cognitive Impairment)

Keyword 1: language
Keyword 2: speech
Keyword 3: technology