Early Classification of Silent Emergency Calls - Exploratory Audio-Based Detection of Pocket-Call-Like Acoustic Patterns

Abstract

Silent calls to the Swedish emergency number 112 place unnecessary strain on emergency-call operations and may delay responses to critical incidents. This thesis investigates whether accidental silent emergency calls can be identified using only caller-side audio. In particular, the work explores the classification of a manually defined subgroup of ”pocket calls”. Two audio feature representations were evaluated: handcrafted acoustic feature representations and pretrained self-supervised BEATs embeddings. Logistic regression and XGBoost classifiers were evaluated across multiple experimental settings, including delayed-speech scenarios, respiratory-audio analysis, and early detection using short audio intervals. The results showed that classification of the complete ”Silent 112” operator-assigned category resulted in poor and unstable performance, while significantly stronger performance was achieved for the ”pocket call” subgroup. Across all experiments, BEATs embeddings outperformed handcrafted features. The strongest performance was achieved using BEATs embeddings together with logistic regression on a 10- second audio window, achieving a recall of 0.92 with no false positives on the heldout test set. Additionally, reliable classification became increasingly feasible after approximately 6 to 8 seconds of available audio. Although the findings are preliminary and limited by the relatively small number of manually identified ”pocket calls”, the results demonstrate the potential of callerside acoustic analysis for supporting automated identification of accidental silent emergency calls. Keywords: