The Emergency Transcriber

The thesis presents a novel situation awareness tool for sensing classification. We proposed a general scheme for sensing, and applied that to build an acoustic tool for teams of first responders and emergency personnel. It constitutes an audio interface for reliably recording and disseminating situation progress as extracted from the team’s audio communications. The tool that we built is intended for emergency teams operating in noisy acoustic environments, where standalone speech recognition systems fail to deliver desired accuracy. Such teams typically follow predefined collaborative workflow as dictated by the relevant engagement protocols, specifying their roles and communications. Given the critical nature of the situation, the vocabulary used is often constrained and dependent on the current stage of the workflow being executed. Treating a traditional speech recognition component as a noisy sensor, the novelty of our tool lies in exploiting knowledge of the workflow to correct the noisy measurements. The intellectual contribution in this exploitation lies in the joint estimation of the current state of the workflow together with the correction of sensed data, given only the noisy (speech) measurements and an overall workflow description. Evaluation shows that the tool provides a significant accuracy enhancement compared to the standalone speech recognition, effectively coping with the noisy environment of emergency teams.