There is usually an audio version available, presumably screen readers are pre-configured to present the audio option on popular captchas by default.

I'm curious now, will have to give it a whirl when I get into the office :D

Suspect Google will rely on their vast knowledge of people's browsing habits based off IP/account/ad-tracking/browser-fingerprinting to skip the user input aspects. Although that said, a screen reader won't have the standard physical interaction clues client-side that a user is a real person, mouse tracking for ex. is probably a moot point. Not really sure how Google will handle those, or if blind users will get a degraded always-on captcha experience.

Either that or a headless screen reader becomes the scraping/botting tool of choice.

Speaking of. Audio captchas are one of the most spooky things I've heard. Noise in the background and muddled voice listing numbers. Think numbers stations, except they try their best to make the numbers unintelligible (just like regular captchas tend to do with written words).

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.

'Good' captchas will have a short audio clip that can be played instead.

Of course not. It's still a pending issue they haven't yet solved.