The emotional state of team members has a critical impact on productivity and software quality, particularly in agile environments where communication is constant and work pressure is common, due to constant deadlines. Given this reality, an approach to continuously and automatically monitor emotional states in a team could be essential to maintaining well-being, preventing burnout, and ensuring sustainable collaboration. However, most prior work on this subject relies on public datasets recorded in controlled environments and non-technical languages, which limits applicability in real-world corporate settings where emotional expressions are subtle and audio quality is often poor. Thus, this paper presents a comparative study of Machine Learning (ML), Deep Learning (DL), and text-based Large Language Models (LLMs) for emotion classi- fication in audio segments from real agile software development meetings. In order to achieve that, we collected and annotated a novel dataset of 782 audio excerpts from four agile meetings (in Portuguese). Our results show that models trained on real meeting audio performed well overall (weighted F1-score around 0.85) but struggled with less frequent emotions, confirming the strong class imbalance in real-world data. DL models such as LSTM and GRU did not outperform simpler machine learning baselines, suggesting that most emotional information was already captured by the audio features. Models trained on acted English data failed to generalize to our spontaneous Portuguese meetings dataset, while text-based LLMs reached similar weighted scores (around 0.8) but still missed subtle emotions, with no clear improvement when dialogue context was added. As such, our contributions include: a novel, domain-specific Portuguese speech emotion recognition dataset, and a systematic comparative analysis of acoustic and semantic (LLM) classification approaches that highlights the challenges of deploying emotion recognition, be it audio or text-based in real-world agile settings.