A non-speech audio CAPTCHA based on acoustic event detection and classification

The completely automated public Turing test to tell computers and humans apart (CAPTCHA) represents an established method to prevent automated abuse of web services. Most websites provide an audio CAPTCHA - in addition to a conventional visual scheme - to facilitate access for a wider range of users. These audio CAPTCHAs are generally based on distorted speech, rendering the task difficult for untrained or non-native listeners, while still being vulnerable against attacks that make use of automatic speech recognition techniques. In this work, we propose a novel and universally usable type of audio CAPTCHA that is solely based on the classification of acoustic sound events. We show that the proposed CAPTCHA leads to satisfactorily high human success rates, while being robust against recently proposed attacks, more than currently available speech-based CAPTCHAs.