What percentage of SpinVox transcriptions require human assistance?

It’s PR 101. Be direct.

But no, SpinVox continues to dig a mountainous hole by not being direct.

Have you been following the catastrophe? Check Alex’s first post and then my follow-up yesterday here and here.

Here’s what SpinVox should have done. The moment the mainstream media picked up on the story, they should have immediately set the record straight.

They should have used their own media — their blog — to manage and direct attention. To run a conversation. Instead they retreated behind closed doors and used their medium simply for publishing a one-way media statement. They should have linked to the BBC post and every other post of those raising issues with the company. They should have then collated the main issues into a Q&A series of challenges with their rebuttals. Central to the rebuttal should have been a point like this:

“SpinVox does use humans to help aid transcription (and these operators can’t access any personal details/anonymisation/etc/etc.). However our brilliant proprietary transcription AI system is extremely reliable. To put this in context, back in the year [whatever] when we launched, our AI system could reliably transcribe approximately 28% of all voicemail messages without human intervention. Fast forward [x] years and as of last month, our AI system processes approximately 97% of messages without any human intervention.”

“Here are some examples of messages that our AI system can process *without* intervention:”

[insert some innocuous examples]

“Here are some examples of messages that, despite the efforts of our genius PHD chaps, aren’t yet able to be fully translated autonomously:”

[insert difficult examples]

I then think SpinVox should give us a process diagram showing exactly how the AI system functions (at ultra high level). A nice powerpoint a la:

1: Receive message
2: Perform initial scan and determine autonomous transcription reliability index [or whatever fancy language they might choose to use]
3: If message contains difficult-to-translate stuff, refer to human operator
4: Piece everything together and deliver to user

Make it clear to everyone how it works. We don’t need IP. We don’t need 100 pages of explanation. But back it up. Explain it with a few charts. Demonstrate the process.

And then, once again, explain that the majority of messages are automated.

Or announce to the world that, for the avoidance of doubt and for clarity’s sake, because 97% of messages are automatically translated, you’re going to put an asterisk in front of any message that’s been human translated.

Really, the problem for most people isn’t that humans are involved in the process — it’s the blurring. The fact that we don’t really know what’s going on.

If you setup an account with a call transcription agency, it’s clear what’s going on. Obviously a human will answer the call and put the answer in text to you.

If you setup an automatic voicemail forwarding company (a la HulloMail), similarly, you know a human isn’t touching any process.

But right now, mainstream media is telling the planet (rightly or wrongly) that SpinVox uses ‘sweatshops’ and that the company was ‘lying’ or being someway disingenuous in regard to their service being entirely/almost entirely operated by a brilliant AI system.

That confusion, doubt and ‘sweatshop’ message is spreading — and, from the conversations I’ve had with normal non-tech (and techie) folk, it’s sticking.

Let’s have some clarity, SpinVox?

Of course… if we don’t get clarity as I’ve described above, it’s going to be painfully clear, surely, that the percentage of human vs machine is shockingly embarrassing. I hope it’s not.