Bypassing Audio reCAPTCHA with Automatic Speech Recognition Models

CAPTCHAs are challenges designed to distinguish humans from automated bots. However, with the growing capabilities of Automatic Speech Recognition (ASR) models, these challenges are increasingly vulnerable to automated resolution. In this paper, we evaluate the feasibility of bypassing the audio versions of CAPTCHAs. We automate the collection and transcription of Google audio CAPTCHAs and compare the performance of several models, including Google Speech-to-text, DeepSpeeach, Whisper, Azure AI Speech, and Deepgram, focusing on accuracy, speed, and cost. The performance results highlight how easily audio CAPTCHAs can be bypassed, in one second for the fastest methods. We also discuss possible countermeasures that could be deployed. We make all collected audio files and results available to the community at: https://gitlab.com/lepolodiou/data-set-bypass-audio-captcha.