Who Pressed First? Deep, Classical and Template-Based Classifiers for Near-Simultaneous Sound Events from a Single Microphone

Competitive games with robots or interactive systems often need to decide which player pressed a buzzer first. With a single microphone, this means classifying overlapping sound events whose onsets may differ by only tens of milliseconds. We compare five classifiers for this task: a convolutional recurrent neural network with attention (CRNN), a Transformer, a Random Forest, a support vector machine (SVM) and a training-free frequency-domain cross-correlation (CC) method that matches Mel-spectrograms against class templates. A dataset of 600 single presses from four electronic buzzers and 1,950 synthetic overlaps (offsets 0–0.5 s) was recorded in a 35 m² room. Each method was evaluated on accuracy for single, near-simultaneous and simultaneous presses, robustness to offset and noise, end-to-end latency, synthetic-to-real transfer and feature representation. The Random Forest achieved the highest mean accuracy (99.3%) with 20.3 ms latency. The CRNN and Transformer were near-perfect on single sounds but reached only 55.5% and 84.2% on simultaneous presses, with latencies above 190 ms. CC ranked third (90.0%) but needed no training, ran in 23.4 ms, stayed at about 0.97 on single sounds at every noise level tested, and transferred best from synthetic to real recordings (0.91, against 0.33–0.66 for the learned models). Under matched real data, the learned models were slightly ahead (CRNN and Random Forest 0.98, CC 0.95). No method dominated. For structured, reproducible sounds deployed in changing rooms, template correlation offers the best balance of speed, robustness and data efficiency. Learned models are preferable where in-room training data are available or sounds overlap heavily in noise.