Emotion recognition through artificial intelligence (AI) is a crucial field of research that can be applied in various areas due to its capacity to interpret and respond to human affective states. In this regard, this study proposes to conduct a documentary review on recent advances in emotion recognition techniques in speech, with a particular focus on Long Short-Term Memory (LSTM) networks. These models are well-suited for processing sequential data, making them highly effective in capturing temporal emotional cues in speech signals. Different scenarios and current challenges have been examined, such as cultural interpretation, model accuracy, ethical implications, and privacy concerns. One of the main issues involves the generalizability of models across diverse populations, as emotional expression can vary significantly between cultures. Additionally, ensuring high accuracy while maintaining computational efficiency remains a priority in system development. Despite these challenges, emotion recognition with AI offers significant potential in areas such as healthcare, for early diagnosis and emotional support, human-computer interaction (HCI), education and marketing, where understanding consumer emotions can improve user experience and engagement. Furthermore, the integration of multimodal approaches, combining speech with facial expressions or physiological signals, has shown promise in improving recognition performance and robustness. As AI continues to evolve, the field of emotion recognition is expected to become increasingly sophisticated, enabling more natural and empathetic interactions between humans and intelligent systems.