The accessibility and ease of use of tools based on generative adversarial networks (GANs) have enabled the creation of synthetic audios or deepfakes. Which can be partially or completely fabricated, with the aim of deceiving, spreading misinformation, producing revenge pornography, and other crimes. This study explores the phonetic similarities and differences between a natural voice and an artificial voice. We apply two methods of generating deepfakes. The first is generated from text to speech, and the second from voice to voice. Then, we perform the analysis of the deepfakes applying the digital forensic methodology UNE 71506:2013 and supported by experimental tests of the digital evidence. The results proved that the automatic programs have a high convergence between the natural voice and the artificial voice. However, the forensic acoustics and phonetic analysis techniques proved differences in the production of certain sounds, such as intonation. Although the samples prove a high similarity in the biometric field, the artificial voice still does not reach a complete equivalence with its natural counterpart in the phonetic aspect. Therefore, our study reveals a clear need to investigate the detection of audio deepfakes and expose it as a source of scientific evidence for the judicial system.