Strong Label Generation for Preparing Speech Data in Military Applications Using CTC Loss

For most military speech datasets, only “weak” textual annotations are available meaning that the content of the speech is known but the exact on- and offset of each word is not. This is primarily due to the high costs associated with manually determining the start and end times of individual words. In military contexts, determining them is especially difficult because of degraded audio quality caused by noise, transmission channels as well as stressed and non-cooperative speakers. However, certain tasks, such as training or evaluating a keyword spotting system, require isolated keywords and thus precise locations of words. This work utilizes the connectionist temporal classification loss of a trained automatic speech recognition system to automatically generate strong labels. Our findings indicate that this approach attains performance levels comparable to systems explicitly designed for converting weak labels into strong ones, and even surpasses them when operating in noisy acoustic conditions such as radio communications, which frequently arise in military settings. Therefore, the presented approach can be used as a semi-automatic annotation tool for preparing military speech datasets because data only has to be weakly labeled by hand. This paper was originally presented at the NATO Science and Technology Organization Symposium (ICMCIS) organized by the Information Systems Technology (IST) Panel, IST-205-RSY ˆ00AC“ the ICMCIS, held in Koblenz, Germany, 23-24 April 2024.