Background: Changes in the speech and language of patients with Alzheimer’s disease (AD) have been reported. Using machine learning to characterize these irregularities may contribute to the early, non-invasive diagnosis of AD. Methods: We conducted cognitive function assessments, including the Mini-Mental State Examination, with 83 patients with AD and 75 healthy elderly participants, and recorded pre- and post-assessment conversations to evaluate participants’ speech. We analyzed the characteristics of the spectrum, intensity, fundamental frequency, and minute temporal variation (∆) of the intensity and fundamental frequency of the speech and compared them between patients with AD and healthy participants. Additionally, we evaluated the performance of the speech features that differed between the two groups as single explanatory variables. Results: We found significant differences in almost all elements of the speech spectrum between the two groups. Regarding the intensity, we found significant differences in all the factors except for the standard deviation between the two groups. In the performance evaluation, the areas under the curve revealed by logistic regression analysis were higher for the center of gravity (0.908 ± 0.036), mean skewness (0.904 ± 0.023), kurtosis (0.932 ± 0.023), and standard deviation (0.977 ± 0.012) of the spectra. Conclusions: This study used machine learning to reveal speech features of patients diagnosed with AD in comparison with healthy elderly people. Significant differences were found between the two groups in all components of the spectrum, paving the way for early non-invasive diagnosis of AD in the future.