The Major Depressive Disorder (MDD) continues to be one of the issues that impair the validity of diagnoses since it relies majorly on subjective clinical interviewing and self-reported inventories. The increasing demands of objective and scalable biomarkers have focused the study on voice analysis that has been identified to be a field that is closely related to the emotional and cognitive conditions. But the recent researches have shown that depression causes a measurable alteration in the acoustic features such as, reduced variability in pitch, slowed speech, extended pause and changed vocal intensity as an outcome of the neurophysiological and motoric disequilibrium in the depressed patients. The current development of machine learning, especially the deep learning and transfer learning algorithms have allowed the potential extraction and exploitation of sensitive aspects in voice such as depression screening and severity with a diagnostic accuracy of up to 85% in constrained environments. Also longitudinal studies have shown that the reduction of depressive symptoms is associated with a measurable recovery in acoustics parameters with time. Nevertheless, its translation into clinical use is limited by the fact that the datasets used are not homogeneous in terms of their characteristics, there is cross-population generalizability, and the ethical aspect is associated with privacy issues. Overall, the use of machine learning in voice is a significant step in the right direction of objective, readily available, and relentless mental health measurement, yet further standardization, validity, and collaboration with other disciplines are required to bring it to life.