Mild cognitive impairment (MCI) is a clinical condition characterized by cognitive decline exceeding normal aging, while not yet reaching the level of dementia. In some patients, it may represent a prodromal stage of a neurodegenerative disease. Since speech changes can occur already in the early stages of MCI, these changes can be monitored and used for non-invasive detection of disorder progression. The motivation of this thesis is to explore the possibilities of acoustic and linguistic analysis for differentiating between different types of mild cognitive impairment. The aim of the thesis is to identify key differences in speech and language manifestations among healthy individuals, patients with amnestic MCI, patients with MCI with Lewy bodies, and patients with Parkinson’s disease with MCI. The analysis used a database containing recordings of spontaneous speech and text reading from healthy individuals and patients diagnosed with MCI. Acoustic and linguistic parameters were extracted from the recordings and subsequently used in statistical and classification analyses. The analysis revealed differences between healthy individuals and patients affected by cognitive impairment, both in acoustic properties and in linguistic aspects of speech production. The classification part included testing machine learning models based on manually selected biomarkers, features selected using statistical testing, and their extension with audio and text embeddings. The results showed that the most important speech biomarkers included mainly parameters related to pauses in speech. Some biomarkers showed a significant relationship with clinical scales and covariates, suggesting a possible association with the clinical condition of the subjects. The classification analysis showed that the best results were achieved mainly in the diagnosis of PD-MCI, while the differentiation of the other groups was less reliable. The combination of acoustic, linguistic, and embedding-based features provided some discriminatory potential; however, the results were not robust enough for reliable automatic differentiation. The thesis therefore highlights the potential of speech biomarkers in the analysis of MCI and the need for larger and more balanced datasets.