Additive manufacturing (AM) of dense metals is prone to pores and cracks that degrade structural performance, making real-time defect monitoring essential. We introduce a lightweight, machine-learning pipeline for in situ crack detection from acoustic emission (AE) during laser powder bed fusion (LPBF). High-frequency AE signals are converted to time–frequency representations, and an autoencoder learns compact embeddings of individual emission events that expose latent structure separating crack signals from background noise. Unsupervised clustering identifies crack-related events, which are then used to train a minimal supervised classifier for real-time inference. Among tested representations, spectrogram-based embeddings provided the most reliable separation of crack and noise events. A simple neural network with just over 100 parameters achieved near-perfect classification accuracy while meeting strict latency constraints. The proposed approach enables robust, low-overhead crack monitoring suitable for deployment in industrial LPBF systems.