Automatic Reliability Estimation for Speech Audio Surveillance Recordings

Being able to monitor communications through environmental recordings is an important asset for a forensic investigator, e.g., to prevent terrorist attacks. On one hand, this is becoming easier thanks to the availability of cheaper and smaller audio recordings devices. On the other hand, the automatic analysis of huge corpora of recording is still far from being an easy task. In this paper we propose a method to analyze speech audio recordings to establish how reliable they are in terms of automatic transcription capability. This can be used to automatically select relevant non-corrupted portions from huge corpora of recordings for analysts to focus on. This can also be used to help an investigator getting a quick feedback about the quality of his / her recording while deploying a system in a noisy environment. The proposed solution is based on a data-driven approach that is computationally cheap and can thus be used to process large datasets.