Google Wants Your Voice Data
00_NOP writes "Peter Norvig, Google's director of research, has told New Scientist that one of the reasons the search engine launched Google Voice is that it needs more human voice data to perfect the sort of 'big data, simple algorithm' probabilistic approach to translating voices to text that drives Google Translate. Norvig says that no one is listening to your calls on Google Voice — it is simply their servers trying to get the translation right."
I will say that the translation of my voice mails is terrible. Although, how can you tell if it is translated correctly if you don't listen to it? You can look for proper English, but even some of my translations are proper English yet still incorrect. (names, etc come out wrong.) Though most of the time it it's just a jumbled mess that I can't deduce the actual meaning of.
How do servers assess whether they've got the translation correct without having a human-in-the-loop to listen to the conversation and concurrently read what the server translated? Maybe the data is anonymous by the time it gets to a human, but it seems like humans need to interface with the voice data somehow to validate that the server is translating accurately.
>simply their servers trying to the translation right
>trying to the translation right
>the translation right
Nicely done.
oh shit! Google accidentally my voice data!
I can tell that they want the voice data badly. They make it very difficult to delete call and voicemail history. You can't delete more than 10 records at a time and even then they go into trash and keep piling up over there. You can delete the data from trash but again only 10 at a time. There is no option to empty the trash. Their help section says that the history is purged from trash after 30 days automatically but only that it isn't. My call history sits in the trash indefinitely unless I painstakingly delete all history 10 records at a time.
They have another server that checks the first server's translation. Part of their work is checking that server's affectiveness, too.
This is the price of "free" services.
I gave up trying to get voice software to work over a decade ago. The reason is that I'm trilingual and use all 3 daily. So the software needs to be able to:
- understand a lousy accent: there are some words I cannot and will never be able to pronounce 'right'
- recognize what language is being spoken (having those 3 and only those 3 preset in the options)
Now I haven't tried Google Voice, but none of the software I've tried or heard about could even remotely do those two basic things.
Non-Linux Penguins ?
it is simply their servers trying to the translation right.
Time to offend someone
And you're surprised why? All voice apps. do this. Always have, always will, and until it's perfected, and we're a long, long way from perfecting it.
Steven
"it is simply their servers trying to the translation right"
I think you a word in your sentence.
"Peter Norvig, Google's director of research, has told New Scientist that one of the reasons the search engine launched Google Voice is that it needs more human voice data to perfect the sort of 'big data, simple algorithm' probabilistic approach to translating voices to text that drives Google Translate. Norvig says that no one is listening to your calls on Google Voice — it is simply their servers trying to the translation right."
I think Google Voice translated the last part of that sentence.
The translation is off pretty far most of the time for my voicemail. But they do end up to be entertaining. Here are two actual translations from google voice:
1) Okay, I don't know why it takes for ever, for your voicemail to pick up. But anyway, I was just calling to tell you that we forgot to while. I will and I told Mrs. Smith and this is best but she signed about it, so I'm gonna shout in the car and have it for her after I pick her up, bye. You Get Out virtual slot is not with us. So, wish me luck. I don't have to run out over there, bye.
2) Hey, it's me. Yeah, I would you like. The has the can't seeing tomorrow night and then there's the car was up for a late night should keep talking about it I'm sure wondering if we could let him get the cantinas left for due to Cardiff blues. I thought I'd run it by you and see what you thought, let me know if any a message or something, so I can out and I will talk to you later. Bye.
Right. Everyone get on Google Voice with funny unnatural accents, unusual intonation and non-native grammar! Let's skew their data.
Comment removed based on user account deletion
There was a Userfriendly.org strip years ago which pretty much summarizes my experience with voice recognition software for the past 15 years. . .
I can't find the link to the comic anymore, but basically, one of the guys in the office had been trying to use voice recog software. Some of his coworkers come to his office. He's not there, but on the screen, they wonder about the mysterious message, "Cod Am Pizza Ship".
The Strip.
Y'all kin have mah voce data. Sheeeit! I warn't doin' nutin' wid it anyhows.
Have gnu, will travel.
I'd be willing to let this happen if google then released the derived heuristics as free open source software. I'll share if you share.
Google is knot an evil umpire. They our hear 2 us with wheel whirled problems. Please stop bash tag google. All your words belong to us.
The next site to slashdot will be ready soon, but subscribers can beat the rush and start slashdotting it early!
So.. They are trying to teach GALaDOS (Google Artificial Lifeform and Disk Operative System) to speak?
This is the same as how they put "Closed Captions" on youtube videos.
Google has no interest in crowd-sourcing the translation or transcription of speech, they want it all automated.
Which is why YouTube Closed Captions SUCK!
Cwm, fjord-bank glyphs vext quiz
In Virginia: "Virginia's wiretapping law is a "one-party consent" law. Virginia makes it a crime to intercept or record any "wire, oral, or electronic communication" unless one party to the conversation consents. Virginia Code 19.2-62. Therefore, if you operate in Virginia, you may record a conversation or phone call if you are a party to the conversation or you get permission from one party to the conversation in advance. That said, if you intend to record conversations involving people located in more than one state, you should play it safe and get the consent of all parties." Google is breaking the law because they are not informing both parties of the recording in advance. I'm sure that all this information is buried somewhere in the terms of service that you click "yes" to without reading it. But, unfortunately for google, that does not satisfy informed consent laws. Some states even require consent on a call by call basis, so blanket consent is automatically invalid and illegal.
$2 says your "consent" is buried in the TOS.
Are you sure that signing up for Google Voice doesn't include a clause giving them permission to store the audio? IANAL, but I would also presume that they could successfully claim that it is obvious that a voice mail service must record the audio in order to store and reproduce it later.
--Udo.
It'd be kind of hard to leave a voice mail if you didn't want the receiving party to record it. Implied consent much?
GV has had an opt-in feature to basically donate each GV transcription to Google with an indicator of whether you thought it was a good or bad transcription. Is there any evidence that Google is delving into your voice data without your consent? Were you expecting a GV transcription *without* a machine at least analyzing the voice data that came in and then discarding it?
Just grab audio from thousands of dialogs or talks on YouTube and test it out.
Except for ending slavery, the Nazis, communism, & securing American independence, war has never solved anything.
Do you not understand what voicemail is? How can record a message for someone without consenting to it being recorded?
We hope your rules and wisdom choke you / Now we are one in everlasting peace
There will only be one voice operating System. Google wants to get there first.
There are many ways to do speech recognition, but one thing they all rely on is training. My guess is that you have a reasonably "average" sounding voice, and therefore the training that was done on other people's voices (specifically, the other users of the app you mention) works well enough for you. For someone with a substantially different sounding voice, it would probably work far worse than Google Voice.
I should also point out that checkbox or no, the app you're using certainly is training. It may not be telling you that it is, but with every utterance you make, it most likely is. If it's not, then it trained on someone else's speech and you just got lucky that it works. That's how speech recognition works, and if you don't believe me, I can direct you to a number of papers on the topic.
I'm sure Google wants to be able to identify people by their voices... I mean in the digital era, where you have pseudonyms, multiple identities, and where portable (micro)phones are proliferating, it would be a mistake not to take advantage of the opportunity to identify or disambiguate people's identity thanks to their voice signature... I'm going to choose my future phone operating systems very carefully...
I love the idea of the feature -- I hate stopping to listen to voice mails. ...but I got the most ludicrous / hilarious translation yesterday. Pure poetry!
"Hi, My name is The bring the Anderson and I was interested in ordering. I'll call. Sarah, Mrs. Kate, on the Hudson birthday. He really liked all here in the for you. So, anyway, I would see it for next Sunday. I'm not sure where they said of. It's Hey Lady, Thank tonight anyway. If you can just give me a back. My phone number is 972."
Good to know about the "helpful? / donate" feature (from a comment above), which I did for this one...
Sure for the person owning the phone, but that hardly is going to provide consent for the person calling you in states that require 2 party consent.