Apple and Google temporarily stop listening to Siri and OK Google queries

Apple will eventually resume human reviews, but only if customers opt in.

Read the whole story

 

[Omega_Prime]

Bug your own house/person and then complain about it?

Services have a cost, in this case privacy.

 

[samanime]

While Apple and Google (and Amazon et al.) definitely shouldn't be holding on to these recordings indefinitely, I do, as a developer, understand the need for humans to review these. They need to periodically listen to what was recorded and then what their service translated it as. That's the only way to ensure it is doing what it is supposed to and make it better.

People who are shocked that a human might listen to their recording are either very tech illiterate and think technology is just "magic" or they are just fooling themselves. Those that are shocked when it is turned over to the police also don't pay attention to the third-party doctrine.

For the record, I personally have none of these devices for this very reason. I don't want a bug in my house. I'm fine spending 30 seconds to type things in.

 

[Hovemond]

It's insane to me that this wasn't opt in from the start. Offer people a dollar or two of credit to opt in or something and they'd be doing it in droves, without any of the privacy/consent issues

 

[ryanja]

Is there any news on a COTS AI program that allows for none of the nonsense the current options have? Ideally, even one that is stored locally? If not, I imagine that it shouldn't be long before there is. Even if it required having it's own physical server or running on a VM on a blade, an AI assistant program that learned & adjusted to speech patterns (speech-to-text recognition programs have been around since easily the days of Win7) and didn't require any outside contact (outside of obvious queries of things like "what's the weather" or "what time does the movie start") seems like it would be a great offering to companies & consumers.

I'd happily get an AI assistant if I owned it, or even licensed it, as long as it didn't come with the uncountable privacy baggage that Siri, Google, Cortana, Bixby, and the others have.

 

[Deleted member 174040]

Opt-in is the way to go and should have been there from the start.

 

[linuxfan66]

While Apple and Google (and Amazon et al.)

People who are shocked that a human might listen to their recording are either very tech illiterate and think technology is just "magic" or they are just fooling themselves. Those that are shocked when it is turned over to the police also don't pay attention to the third-party doctrine.

without question, having no human reviews would be detrimental to service quality long term.
the fact noone considered that until now is beyond me
:shrug:

 

[Iphtashu Fitz]

While Apple and Google (and Amazon et al.) definitely shouldn't be holding on to these recordings indefinitely, I do, as a developer, understand the need for humans to review these. They need to periodically listen to what was recorded and then what their service translated it as. That's the only way to ensure it is doing what it is supposed to and make it better.

Back in the 90's I worked for a few years for a speech recognition company that was specifically targeting the telephone industry. On a fairly regular basis our speech engineers would distribute lists of phrases and ask us to call a specific number and read them off. They'd collect the recordings and use them for further training of the speech recognition engine.

I wonder if Apple, Google, etc. could incentivize some of their users to help out in a similar regard. For example, the Apple Developer Program costs $99/year. They could potentially do something like offer half price off for the year if you spend, say, two hours throughout the year reading content for Siri's engineers to analyze. It wouldn't be rocket science. Just develop an app that you log into with your developer account, have the app display words/phrases for you to read, then record you saying them & upload to Apple. Once the app determines you've recorded the minimum amount of data then a discount is applied to your developer account, maybe for the next years subscription.

 

[samanime]

Is there any news on a COTS AI program that allows for none of the nonsense the current options have? Ideally, even one that is stored locally? If not, I imagine that it shouldn't be long before there is. Even if it required having it's own physical server or running on a VM on a blade, an AI assistant program that learned & adjusted to speech patterns (speech-to-text recognition programs have been around since easily the days of Win7) and didn't require any outside contact (outside of obvious queries of things like "what's the weather" or "what time does the movie start") seems like it would be a great offering to companies & consumers.

I'd happily get an AI assistant if I owned it, or even licensed it, as long as it didn't come with the uncountable privacy baggage that Siri, Google, Cortana, Bixby, and the others have.

Something like this with absolutely no human listening ever is still decades away. We are getting better and better, but there are 10s of 1000s of accents and some still need a lot of work. Even those that are well understood still need human review periodically to make sure it is working and to find edge cases. But, for a (good) service to never have a human listen just isn't possible yet. You'll need to listen to smaller and smaller samples, but still listen to some.

 

[andygates]

Back in the 90's I worked for a few years for a speech recognition company that was specifically targeting the telephone industry. On a fairly regular basis our speech engineers would distribute lists of phrases and ask us to call a specific number and read them off. They'd collect the recordings and use them for further training of the speech recognition engine.

I wonder if Apple, Google, etc. could incentivize some of their users to help out in a similar regard. For example, the Apple Developer Program costs $99/year. They could potentially do something like offer half price off for the year if you spend, say, two hours throughout the year reading content for Siri's engineers to analyze. It wouldn't be rocket science. Just develop an app that you log into with your developer account, have the app display words/phrases for you to read, then record you saying them & upload to Apple. Once the app determines you've recorded the minimum amount of data then a discount is applied to your developer account, maybe for the next years subscription.

Test cases the engineers can think up are exactly not the edge cases that make the service more useful.

 

[cpragman]

When Siri mangles a request, I’d like there to be some way (a keyword or something) to send the mangled request to Apple for QA.

If all I can do is opt-in globally to their QA, then it won’t be very effective at finding and fixing problems, and it won’t solve the “always listening” problem at-hand today.

 

[Dilbert]

"There will be no networked computer systems on this Battlestar for as long as I'm in command."

Adama was onto something.

Colonists also elected a traitor celebrity populist to presidency and that went pretty badly form them too.

 

[Iphtashu Fitz]

Is there any news on a COTS AI program that allows for none of the nonsense the current options have? Ideally, even one that is stored locally? If not, I imagine that it shouldn't be long before there is. Even if it required having it's own physical server or running on a VM on a blade, an AI assistant program that learned & adjusted to speech patterns (speech-to-text recognition programs have been around since easily the days of Win7) and didn't require any outside contact (outside of obvious queries of things like "what's the weather" or "what time does the movie start") seems like it would be a great offering to companies & consumers.

I'd happily get an AI assistant if I owned it, or even licensed it, as long as it didn't come with the uncountable privacy baggage that Siri, Google, Cortana, Bixby, and the others have.

Something like this with absolutely no human listening ever is still decades away. We are getting better and better, but there are 10s of 1000s of accents and some still need a lot of work. Even those that are well understood still need human review periodically to make sure it is working and to find edge cases. But, for a (good) service to never have a human listen just isn't possible yet. You'll need to listen to smaller and smaller samples, but still listen to some.

Yup, those edge cases are killers. My wife and I occasionally take care of a neighbors dog named "Z" (yup, that's her full name). Any time we rely on Siri to dictate text messages, etc. about her we get a wide range of terrible responses. Just consider the difficulty of a computer trying to understand nicknames, much less unusual proper names, etc.

 

[Fred Duck]

VRT NWS also said that 153 of the 1,000 recordings it listened to "were conversations that should never have been recorded and during which the command 'OK Google' was clearly not given."

Wait, I thought "OK google" was the signal to start recording/listening. How could it ever have been clearly given in a recording, unless it's always recording? Or am I misunderstanding this?

 

[StuartDole]

Well, Apple Maps has a "report an issue" option - just include something like that for Siri. We have a macaw named "Sargie", and our typical morning "Hi Sargie!" gets all the Siris in audio range all in a knot!

 

[droidmonkey]

Headline is confusing, I thought there was a global shutoff of the devices themselves.

 

[MacRat]

Wait, I thought "OK google" was the signal to start recording/listening. How could it ever have been clearly given in a recording, unless it's always recording? Or am I misunderstanding this?

The system has to be listening in order to hear "OK Google" or "Hey Siri"

 

[Fatesrider]

Is there any news on a COTS AI program that allows for none of the nonsense the current options have? Ideally, even one that is stored locally? If not, I imagine that it shouldn't be long before there is. Even if it required having it's own physical server or running on a VM on a blade, an AI assistant program that learned & adjusted to speech patterns (speech-to-text recognition programs have been around since easily the days of Win7) and didn't require any outside contact (outside of obvious queries of things like "what's the weather" or "what time does the movie start") seems like it would be a great offering to companies & consumers.

I'd happily get an AI assistant if I owned it, or even licensed it, as long as it didn't come with the uncountable privacy baggage that Siri, Google, Cortana, Bixby, and the others have.

A local-only brain in the box AI assistant is fine, and with some of the new chips and such that will aid in cross-language AI programming, one might be developed (I'll be looking into that myself).

But requiring some Internet connection for basic functionality (like voice to text services in composing, turning on and off IoT devices, adjusting temperatures, etc.) that don't actually NEED an Internet connection to do (just LAN access) is creating the same problem of putting your information out there. Whether or not someone listens to it will always be a question. After all, it's OUT THERE, and there's no absolute certainties about what happens to the data that goes out there or who can get to it.

At least with a wholly local set-up, you have a much better chance to monitor access to your "brain" than you have if your "brain" is outsourced to somewhere in the etherspace.

Oh, and for the record, those voice to text programs (like Dragon Naturally Speaking) have never achieved a level of detection that is faster than I can type. Its accuracy never seems to get above 98%, which some folks may think is fine, but the fact is, a 2% error rate drops the actual output of text by about 25% just to go back and fix the errors. It also ruins the continuity of thought when something was misinterpreted and one tries to correct it.

Being an author, I've tried these things before, and they never do what I need them to do reliably enough to replace typing. Unless I speak at a frustratingly slow place, the program just can't do the job, and even then, the error rate remains unchanged (it only increases the faster I talk, and I'm very good at annunciation - it's not very good at hearing what's being said).

A lot of news services use those kinds of things to create a transcript of a video, and that's usually unintelligible garbage often with hilarious results.

An AI would probably do better, because it has a faster learning curve.

 

[peachpuff]

Opt-in is the way to go and should have been there from the start.

Don't think you understand how these companies work, it's easier to opt everyone in, then apologize when they're caught with their pants down and then offer an opt out.

 

[Fred Duck]

The system has to be listening in order to hear "OK Google" or "Hey Siri"

Hearing ≠ listening.

 

[adespoton]

It's insane to me that this wasn't opt in from the start. Offer people a dollar or two of credit to opt in or something and they'd be doing it in droves, without any of the privacy/consent issues

Just offer them a chance to win a free app in the app store and they'll opt in in droves.

 

[Pen Drive]

These "services" are rubbish and add nothing to the value of devices. They were never demanded by customers. They have been forced on users by tech companies.
At the very least let customers remove Siri, Cortana, Alexis or whatever but you don't let us do that. Its not cool its stupid.
Stop spying on your customers. Stop messing with your customers. Your rubbish AI technology isn't wanted except by possibly disabled users or users who cannot read or write.
Just what are you playing at?

 

[ZippyPeanut]

These "services" are rubbish and add nothing to the value of devices. They were never demanded by customers. They have been forced on users by tech companies.
At the very least let customers remove Siri, Cortana, Alexis or whatever but you don't let us do that. Its not cool its stupid.
Stop spying on your customers. Stop messing with your customers. Your rubbish AI technology isn't wanted except by possibly disabled users or users who cannot read or write.
Just what are you playing at?

These services offer a lot of value to the devices.

 

[Dilbert]

It's insane to me that this wasn't opt in from the start. Offer people a dollar or two of credit to opt in or something and they'd be doing it in droves, without any of the privacy/consent issues

Just offer them a chance to win a free app in the app store and they'll opt in in droves.

Yes. There's an hour long line out the door and down the street whenever Ben And Jerrys has a free ice cream day. Single scoop. Worth maybe $3. This is an extremely affluent area where they are no houses worth under 1 mil and many worth a lot more than that, and everyone drives a Tesla. But they will waste an hour to get one free ice cream.

 

[DarKnyht]

So, this would be why prefer to require a button be pushed to activate the service. It is a tradeoff, but a useful one for when I am driving a vehicle or otherwise in a situation where stopping to request information (read a text, what's the weather, etc.) or input information (typing a replay, picking a song, etc.) would be detrimental.

 

[RareHamster]

When Siri mangles a request, I’d like there to be some way (a keyword or something) to send the mangled request to Apple for QA.

If all I can do is opt-in globally to their QA, then it won’t be very effective at finding and fixing problems, and it won’t solve the “always listening” problem at-hand today.

I’ve always assumed that the “tap-to-edit” button (on the Siri screen displaying the transcribed speech) submits the recording and user edits to some kind of internal QA process, whether automated or human.

 

[albopf]

VentureBeat said, “Apple and Google halt human voice-data reviews over privacy backlash, but transparency is the real issue”. Will we ever “see transparency”? (Yes, I know, reductio ad absurdum /s)

 

[graylshaped]

[the usual Apple hate]

Zippy, do your reputation a favor.

If "Apple" is part of the story, just walk away. Walk away.

 

[graylshaped]

Opt-in is the way to go and should have been there from the start.

I find it distasteful to agree with you.

 

[Voyna i Mor]

While Apple and Google (and Amazon et al.) definitely shouldn't be holding on to these recordings indefinitely, I do, as a developer, understand the need for humans to review these. They need to periodically listen to what was recorded and then what their service translated it as. That's the only way to ensure it is doing what it is supposed to and make it better.

Back in the 90's I worked for a few years for a speech recognition company that was specifically targeting the telephone industry. On a fairly regular basis our speech engineers would distribute lists of phrases and ask us to call a specific number and read them off. They'd collect the recordings and use them for further training of the speech recognition engine.

I wonder if Apple, Google, etc. could incentivize some of their users to help out in a similar regard. For example, the Apple Developer Program costs $99/year. They could potentially do something like offer half price off for the year if you spend, say, two hours throughout the year reading content for Siri's engineers to analyze. It wouldn't be rocket science. Just develop an app that you log into with your developer account, have the app display words/phrases for you to read, then record you saying them & upload to Apple. Once the app determines you've recorded the minimum amount of data then a discount is applied to your developer account, maybe for the next years subscription.

Won't work, because people reading set phrases for money will speak differently from the general public.

 

[Voyna i Mor]

VentureBeat said, “Apple and Google halt human voice-data reviews over privacy backlash, but transparency is the real issue”. Will we ever “see transparency”? (Yes, I know, reductio ad absurdum /s)

Well, of course you can. It's on a scale and you can easily estimate visually the transparency of a substance.
In the limit if you put alumina next to glass of the same thickness you can just see that alumina is only about 99% transmissive.

 

[shryke1441]

If you actually GO to google and try and turn off the Voice & Audio, you are presented with this message:

Pausing Voice & Audio Activity may limit or disable more personalized experiences across Google services. For example, Google may not understand you when you say "Hey Google" to speak to your Assistant.


This seems to contradict their 'message from a few months ago', where Voice & Audio isn't necessary to use the assistant...

 

[ewelch]

Why don't they have computers transcribe such voice recordings and let the reviewers read them first. If there's a question about the transcrxiption, only then does someone listen. And the voices can be distorted so they are not identifiable.

Of course, there's only so much you can do to anonymize phrases such as, "I'm a stable genius," or "I'm the least racist person in the world." But then that person is likely not going to be embarrassed anyway if those recordings leak.

 

[thetay48]

It’s interesting how Amazon, Apple, and Google have handled these privacy problems...

Apple gets caught and issues a statement on how they were doing things and paused the program (voluntarily) in just over a week. And say they’ll try for more privacy and so on

Google gets caught, issues a similar statement on how they’ve done it and but only after Germany goes after them to stop recording, they pause their program and this happened within a month

Amazon gets caught big time, issued some vague PR statements, do some damage control, say they’ll look into how recordings are done and try to listen less, but continue their program and hasn’t changed much, if anything, since this was reported in April

 

[klarg]

There should be a law (laws) forbidding any and all data gathering without express opt-in, with annual re-notification of any opt-ins, and a straightforward means to immediately cancel any such permissions. As a starting point.

 

[albopf]

VentureBeat said, “Apple and Google halt human voice-data reviews over privacy backlash, but transparency is the real issue”. Will we ever “see transparency”? (Yes, I know, reductio ad absurdum /s)

Well, of course you can. It's on a scale and you can easily estimate visually the transparency of a substance.
In the limit if you put alumina next to glass of the same thickness you can just see that alumina is only about 99% transmissive.

I was thinking more of the Cheshire Cat, “I can see nothing,” and Alice, “My, you must have good eyes.”

 

[dani_sht]

While Apple and Google (and Amazon et al.) definitely shouldn't be holding on to these recordings indefinitely, I do, as a developer, understand the need for humans to review these. They need to periodically listen to what was recorded and then what their service translated it as. That's the only way to ensure it is doing what it is supposed to and make it better.

People who are shocked that a human might listen to their recording are either very tech illiterate and think technology is just "magic" or they are just fooling themselves. Those that are shocked when it is turned over to the police also don't pay attention to the third-party doctrine.

For the record, I personally have none of these devices for this very reason. I don't want a bug in my house. I'm fine spending 30 seconds to type things in.

Thank you. I was about to write a similar rant.
Some of the services that those tech giants offer, I find really convenient. So I would gladly donate some data for the machine learning algorithm to get better and offer even better service in the future. I don't really get people's problems with allowing their voice recordings being heard by strangers which are there to do a job. They do go through many recordings and I believe the data is anonymous and that makes your recording nothing more than statistical data.
And for the record, if you concerned about your recordings reaching the police, you are at the very least being an asshole behind closed doors.

 

[GodelEscherDali]

The system has to be listening in order to hear "OK Google" or "Hey Siri"

Hearing ≠ listening.

More to the point: listening ≠ recording

 

[Elsmarc]

These "services" are rubbish and add nothing to the value of devices. They were never demanded by customers. They have been forced on users by tech companies.
At the very least let customers remove Siri, Cortana, Alexis or whatever but you don't let us do that. Its not cool its stupid.
Stop spying on your customers. Stop messing with your customers. Your rubbish AI technology isn't wanted except by possibly disabled users or users who cannot read or write.
Just what are you playing at?

No one has to have Siri turned on. Do you not know Siri (and I assume the G equivalent) can be turned off? I remember when I bought the iPhone I have Siri was off by default. I had to go to Settings to turn it on and train it to my voice.

No one is forcing you to use Siri. My car has lots of buttons on the steering wheel, few of which I have ever used. The company has not forced me to use them. They are included for the people who DO want to have those options available.

I may not technically be able to "remove" Siri software from the phone, but turning it off accomplishes nearly the same thing.

Personally I rarely use Siri, but from time to time I do and I like it. And I personally could care less if the company uses voice snippets for improvement. I think many people are totally anal about some trivial things.