Apple employees are listening to Siri conversations | Hacker News Reader

Apple employees are listening to Siri conversations

reddit.com

13 pointstomlagier3 comments

I posted this not to bash on Apple, but to highlight that this is not just a Google issue - any voice recognition will be trained on labeled data, and that labeled data will largely come from users.

This has happened for years, across anyone with a voice recognition model, and will continue to happen until the market or regulations say otherwise.

We can, and should, debate the ethics of voice recognition models in general and the practices around collection and handling of that data. I just want to move the discussion beyond "Google is evil" and towards "this is a cornerstone of building voice recognition software - how do we deal with that?"

> "this is a cornerstone of building voice recognition software - how do we deal with that?"

Maybe by using paid employers or volunteers who explicitly agree their voice to be processed for machine learning and voice recognition improvement.

Something like Common Voice¹ by Mozilla?

¹https://voice.mozilla.org/

That's a really cool project, props to Mozilla for spearheading it. I'm especially a fan of the clear callout about what profile data is public.

My thoughts on why a public dataset like this doesn't seem to have been adopted by the other BigCo's

- $$$: having a better voice recognition model is moat. I think this is the most obvious answer, but I think there are likely other factors as well.

- Representation: perhaps users who contribute voice recordings to an open project do not provide a complete model of people who use voice recognition software, leading to tough-to-bridge gaps (especially across languages - seems this project is English-only)

- Size: I'm not a ML person so I don't have any idea the amount of data needed to accurately train a good voice recognition model, but I'd guess that 2.3k hours of validated content isn't enough to really do the trick. Maybe it's difficult to get people to volunteer for the quantity necessary to build a good model.

- Quality: Getting representative data might be harder when it's via explicit donation rather than identified error cases.

- Privacy: I actually think that having the training data available to the public presents an interesting set of privacy challenges. It's available through explicit consent, so that's a big benefit. However, it hugely amplifies the possible exposure of a person's voice which means that you need to have strict controls around the content of the recordings. You also need to be careful about rendering any particular donator vulnerable (could you hijack someone's voice and use it to brute-force other voice-matched assistants, for example?)

I'm curious whether, at the end of the day, it's practical to build a high-quality model from only explicitly donated clips.

Here's something I'm wondering about now as well. In many states in the US, it's legal to record people without their consent in public places. Could there be some sort of public voice data collection, sort of like street view? That would be completely anonymous, and might give more representative data, though I'm sure external noise would be a big challenge.

Reply on news.ycombinator.com