Development of a Neural Network Algorithm to Detect Soldier Load from Environmental Speech

The objective was to develop a model based on speech input that can identify when team members need adaptive autonomous assistance. Human teams often adjust their behavior to work cohesively and effectively as a team. Similarly, it is beneficial for autonomous agents to be able to adaptively adjust to team needs. We constructed a convolutional recurrent neural network model based on those developed for the recognition of emotion from speech. Audio recordings from a recent field exercise were used to train and validate the model. These data were labeled according to whether the speech occurred during an engagement (engaged, neutral, or no-speech). The model classified more than 99% of the training, validation, and test sets correctly. This information will allow us to design systems in which autonomous agents can prioritize, assist with, and take autonomous control of tasks.