This study explores the application of Machine Learning (ML) techniques to analyze spontaneous speech (SS) for early detection and prediction of cognitive impairment in Alzheimer's disease (AD). The research aims to leverage paralinguistic features extracted from SS to discriminate between different degrees of cognitive impairment and predict cognitive domain performance. The study involved 1500 participants, including individuals with subjective cognitive decline (SCD), mild cognitive impairment (MCI), and Alzheimer's disease dementia (ADD). Physical-acoustic features were extracted from voice recordings, and several ML models were evaluated using cross-validation to identify individuals without cognitive impairment, those with MCI, and those with ADD. The results showed that ML techniques could effectively distinguish between SCD and ADD (F1 = 0.92) and between SCD and MCI (F1 = 0.84). Additionally, the models exhibited strong correlations (≥0.5) for predicting cognitive domains such as attention, memory, executive functions, language, and visuospatial ability. The study highlights the potential of a brief and cost-effective SS protocol in distinguishing cognitive impairments and forecasting cognitive performance, opening new avenues for developing screening tools and remote disease monitoring.This study explores the application of Machine Learning (ML) techniques to analyze spontaneous speech (SS) for early detection and prediction of cognitive impairment in Alzheimer's disease (AD). The research aims to leverage paralinguistic features extracted from SS to discriminate between different degrees of cognitive impairment and predict cognitive domain performance. The study involved 1500 participants, including individuals with subjective cognitive decline (SCD), mild cognitive impairment (MCI), and Alzheimer's disease dementia (ADD). Physical-acoustic features were extracted from voice recordings, and several ML models were evaluated using cross-validation to identify individuals without cognitive impairment, those with MCI, and those with ADD. The results showed that ML techniques could effectively distinguish between SCD and ADD (F1 = 0.92) and between SCD and MCI (F1 = 0.84). Additionally, the models exhibited strong correlations (≥0.5) for predicting cognitive domains such as attention, memory, executive functions, language, and visuospatial ability. The study highlights the potential of a brief and cost-effective SS protocol in distinguishing cognitive impairments and forecasting cognitive performance, opening new avenues for developing screening tools and remote disease monitoring.