SpinVox is people! It's people! That's the Soylent Green cry of the BBC this morning -- its tech correspondent, Rory Cellan-Jones, is sad to find out that SpinVox is using people to translate his voicemails into text, rather than a super-computer called D2, or 'the Brain'.
The worry is that SpinVox is sending our voicemails overseas to be translated by loose-lipped call centres, who might blab our secrets everywhere.
The BBC reports that a Facebook group for people who worked at a SpinVox call centre in Egypt shows screenshots of whole voicemails that were sent to people to be transcribed, including what looks like confidential business info.
SpinVox says its voice-recognition system relies on the Brain, but uses people when the computer can't recognise some words. In fact, it says this combo of computer and human brain power is the big idea behind SpinVox.
The marketing-speak on SpinVox's Web site makes its system sound like a Big Thought-style black box surrounded by white-coated boffins, but the reality is more technical and mundane.
"The machine component of the system converts as much as it can," SpinVox's Johnathan Simnett told us. "If it doesn't recognise a word or a sound, then it refers that portion of the message to a language expert. Then the machine learns and it shouldn't have to ask again."
SpinVox says its 'language experts' are trained call-centre people around the world, which are necessary for the service to be up and running 24 hours a day. But it insists call-centre employees never see whole voicemails -- just small chunks.
It told us the screenshots and audio on Facebook are training samples, and don't affect anyone's privacy. For example, said Simnett, "there's a bit of audio where somebody's going 'Leave a message!' and somebody in the background's going 'Yeah, I'm leaving a message!'."
The call centre on the Facebook page failed to get hired after training, said Simnett.
We're not in love with the thought of call centre-ites giggling over our naughty voicemails, so how much of our chat is going through their possibly prurient ears? It depends.
Each time you get a voicemail from a new person, that's a new funky accent and vocabulary for the system to learn.
"Right at the beginning, it might be that 100 per cent of that message is being converted by a human being -- but that's not one human being just looking at your messages, that's a bunch of human beings operating around the world," Simnett said.
It insists that the amount of machine conversion is constantly increasing as the machines learn. It wouldn't tell us the total percentage of voicemails that get the human touch, saying it could help competitors guess its technique.
The BBC's Cellan-Jones tested SpinVox's claims by sending the same recording several times to see if the computer would come back with the same transcription. He says in his blog that the results were wildly different.
SpinVox explains this by saying that even if the voice mail is an identical recording, each one is different because of noise or delay on the line -- which can make a recording sound dissimilar to a machine.
As for privacy, SpinVox insists that it sticks to the letter of the law. It stores data within the European Economic Area, as enforced by the UK Data Protection Act, and everything sent to its overseas call centres is encrypted.
"The security we have is the same sort that you'd have
with your bank. When you call your bank, you're giving an operator totally
complete information -- they could write that down. Our guys can't even do
that," Simnett said.
SpinVox is also struggling with cash, offering its employees stock options rather than their monthly paycheques. It admits it's having cash-flow problems, but denies it's because it's blowing all its money on call centres instead of using its fancy computer.
In the end, SpinVox says that even if it wanted to, it couldn't rely on call centres to do all the translating, because it's handling so many calls.
"The idea that we're running a Mechanical Turk operation -- don't look behind the curtain because this amazing machine that translates your speech is just some bloke writing it down -- is nonsense; we couldn't possibly scale," claims Simnet.
We think SpinVox could have avoided this controversy if it had toned down its super-computer marketing message and been clearer that our voicemails will probably be peeked at by overseas call centres, even if it's only in little pieces. Then we wouldn't feel so cheated that there's not a giant robot answering our calls.
SpinVox offered us an account to try it out -- as with all our review samples, provided to us for free -- and we'll be giving it a run-through ourselves in the near future.