Typical seismic waveform datasets comprise from hundreds of thousands to several millions records. Compilation is performed by time-consuming handpicking of phase arrival times, or signal processing algorithms such as cross-correlation. The latter generally underperform compared to handpicking. However, inconsistencies across and within handpicked datasets creates disagreement between observations and interpretation of Earth’s structure. Here, we exploit the pattern recognition capabilities of Convolutional Neural Networks (CNN). Using a large global handpicked dataset, we train a CNN model to identify the seismic shear phase SS. This accelerates, automates, and makes consistent data compilation. The CNN model is then employed to identify precursors to SS generated by mantle discontinuities. The model identifies precursors in stacked and individual seismograms, producing new measurements of the mantle transition zone with quality comparable to handpicked data. The capability to rapidly obtain new, high-quality observations has implications for automation of future seismic tomography inversions and body wave studies.