Hacker News Reader

Comment by ezekg | Hacker News Reader

But Google's speech-to-text is the epitome of machine learning.

Yes, but if you are paying them extra to specifically opt out of users' data being used for that purpose, it's extraordinarily unlikely that they are doing so anyway.

But it's unlikely to train on user inputs/outputs, because they wouldn't know if they were correct

There is this thing called unsupervised learning. It's quite possible to use this data for further fine tuning of models if the confidence outputs of the current models are high enough that the data is properly labeled, even if that labeling was done automatically. This is a quick way to bootstrap a small set of labelled data into a larger one. Whenever errors are detected later on you can correct for those and then retrain.

Data that customers are sending to an ML speech-to-text API seems like exactly the sort of data that one would want to save for training future models. Maybe I'm too cynical, but I have no confidence that Google throws away any data that they can get their hands on.

I mean you can just read their description:

By default, Speech-to-Text does not log customer audio data or transcripts. To help Speech-to-Text to better suit your needs, you can opt into the data logging program. The data logging program allows Google to improve the quality of Speech-to-Text through using customer data to refine its speech recognition service. As a benefit for opting in, you gain access to discounted pricing.

Why is the only factual answer in this thread downvoted?

People love to hate on Google here apparently, inconvenient facts to the contrary aside.

I wouldn't be so quick to take their own description as fact.

If you can't trust anything or anyone in life, that's a hard way to live.

Yeah, they could by lying - but they'd be opening themselves up to lawsuits for questionable gain. It would be a dumb risk/reward calculation. That's not to say companies (or people) don't do dumb things, but let's give them the benefit of the doubt until we have reason to believe otherwise.

Over the last decade, most speech-to-text services admitted keeping and humans listening to source recordings beyond what was disclosed in terms of service.

In fact, one of the early “AI/ML” transcription services (which at the time blew away competitors including Google) turned out to be almost zero machine processing, almost entirely a human transcription farm in Egypt.

With almost 100% of the services, including those provided by members of FAANG, admitting this, it seems the risk/reward calc didn’t deter, and a “give the benefit of the doubt” algo seems inappropriate.

I didn't say anyone or anything, in this case we're talking about a company that data mines as one of their primary business models that has been caught lying about the extent of data collection multiple times. They've also been sued successfully multiple times and have settled cases out of court (i.e., paid settlements privately).

I'm not sure what happened, but there used to be a time where not trusting a company like Google would be the smart stance to take. How could anyone with that much power possibly be trustworthy?

The tides have shifted, it seems. People now default to trusting big tech, trusting big govt, trusting big pharma. It was a lot of work, years of propaganda, but they did it!

If you don't trust them to abide by their contracts and agreements, don't use their services. That seems like a minimum requirement to do business with any party.

No point complaining that other people are willing to do business with them.

They could keep the recordings and have someone verify whether the output is correct.

Reply on news.ycombinator.com