News

Spinvox - will it benefit from exposure story?

by Guy J Kewney | posted on 23 July 2009


You may well be planning to sign up to Spinvox, just to see if its voicemail transcription service actually works, after it got a huge publicity exposure this morning. Trouble is, Spinvox probably won't see it as a publicity boost, because it has been accused of cheating - or worse!

Spinvox uses (it claims) a huge speech-recognition computer. It listens to your voicemails, and converts them to text - and then texts them to you. Or, at least, that's what we all thought, until Rory Cellan-Jones told us that he thinks the work is done by call-centre staff. "You can tell," he explained to BBC Radio 4 listeners, "by sending the same message several times, and getting several different text conversions."

If he's right, Spinvox could be in deep trouble. It would, obviously, be in trouble from customers, who signed up for the voicemail service, and who believed that no human would listen to them - but even more, it could be in financial trouble.

Already, the company has raised doubts about its finances, by offering to pay staff in shares, rather than real money - blaming the credit crunch.

Spinvox says that it captures spoken words "and feeds them into a Voice Message Conversion System, known as ‘D2’ (the Brain), and spits them out as text content." The web site explains:

So D2’s pretty smart. It’s bound to be, as D2’s a combination of artificial intelligence, voice recognition and natural linguistics. But it also knows what it doesn’t know and is able to call on human experts for assistance. It learns all the time about how we speak, and what we say, from the mundane to the ridiculous and so is able to convert what you mean to say.

Not every message; it admits that a proportion have to be handed to humans - but the question is - what proportion?

The company refuses to say. But no geek will be able to resist the test performed by Cellan-Jones - send the same voice message five times, and see if it looks like a computer has done the conversion. If enough techies flood the Spinvox system with this sort of test, staff could be overwhelmed.

There is another problem.  The Information Commissioner's Office (ICO) told the BBC that it has contacted Spinvox "to ensure that its entry on the data protection register is both accurate and complete, especially with regards to the transfer of personal data outside the European Economic Area".

The story arose when former call centre employees started blogging about the work they were doing, and they claim they are doing "the majority of conversions."

As to whether the allegations are accurate or not, is something only Spinvox can know for sure. It's hard to doubt that a large number of voicemails are having to be referred to "experts" because any speech recognition researcher can confirm the difficulty of doing speaker-independent transcription. Even with hi-fi sound, powerful computers struggle with variations on accent, voice tone and other individual differences - over the phone, it can be virtually impossible to get data processing to work correctly.

That a proportion of the voice messages are referred to call centre staff need not be doubted; but it's not obvious that the staff would actually know how many were handled by machine.

Spinvox, which could dispell the doubts with a simple: "Around ten per cent" or "sometimes as much as 40%" respond, has done the worst possible thing from a PR point of view; it has refused to reveal the proportion, saying that it is highly sensitive and confidential.

Inevitably, speech experts will say, this implies that the failure rate is much higher than the company's prospectus has suggested, and this in turn suggests that the company's future is in doubt.


Technorati tags:   
Send reinforcements, we are going to a dance - You can discuss this article on our discussion board.