I'm less aware of how the more modern captchas work under the hood, but I do know that the audio interface (provided for for accessibility) of an older version

I'm less aware of how the more modern captchas work under the hood, but I do know that the audio interface (provided for for accessibility) of an older version of captcha was used as an attack vector for defeating it. My understanding is that the optional audio file was run through an STT program to generate the answer text so that requests could still be automated. Translating speech to text is less computationally expensive than OCR, especially when captchas were typically using hard to read or intentionally distorted images of text. Someone who knows more about this than me can certainly school me on that last bit, however.