A series of recent postings from the BBC have questioned the validity of the SpinVox as a “speech technology company.” The posts (one news item and a blog post) raise questions around whether there’s an actual speech engine powering the SpinVox service or just an elaborate archipelago of call centers flung around the world, transcribing messages by hand, as some have accused.
The BBC cites photos from a Facebook group created by staff at an Egyptian call center, RAYA, which used to work for Spinvox, “containing what appears to be sensitive commercial information” as an indication that the human transcription may play a bigger role than SpinVox would like to let on. The blog especially seems to hint that almost all of SpinVox’s transcriptions are done by hand. The company, for its part, has fired back, calling the BBC’s reporting “incorrect” and “inaccurate,” in its own blog response.
Christina Domecq, SpinVox CEO, goes one step further in her response to the Guardian. Domecq says, “The majority of calls are fully automated.”
I was actually able to find the Facebook group that the BBC mentions, “Sp!nVox R@Y@,” pretty easily. It hasn’t been updated in about a year, which may mitigate its importance since SpinVox claims to be upping the proportion of automation behind its service. That said, among the claims made in the group’s description are: “we work behind the scenes…we are the invisible heroes[sic]… we are Tenzing… we are unknown and they deny our existence[sic] !!”
The photo which BBC alleges might contain sensitive commercial information is pictured below.
“The photo is from a training session – that is a dummy message,” claims Rachael Lyons, director of North American communications for SpinVox, in an email to Speech Tech–in other words, a fake message that isn’t sensitive at all.
“[The RAYA agents] were using training data – a model system that SpinVox uses to evaluate the quality of call centre support before contracting with the supplier to handle real user data,” writes SpinVox on its blog. “The training system will require individuals to convert full messages in order to establish their speed and accuracy.”
“This would not be the case on a live customer system where should the VMCS system need assistance in learning, operators would only be presented with portions of any message for assisted learning,” the blog adds.
To prove the company’s point, Ms. Lyons cites a number of items from the group’s photos:
- Pictures show training screens you can tell that because they have other apps running – doesn’t happen in live systems
- Screens have whole messages on them – doesn’t happen in Live systems
- Pictures show training presentations and materials
- Audio is known training audio we use all the time
- Dates on the site were when SpinVox was testing RAYA
- SpinVox hasn’t required for the site to be taken down because it doesn’t breach security
Subsequent to the trial, SpinVox says that the RAYA call center was not retained to handle live data. It says the same for Kencall in Kenya, which failed to meet criteria and, incidentally, also has a Facebook group that complains about working for SpinVox that also hasn’t seen an update for almost a year.
The BBC’s only named source in the article is Kareem Lucilius who said he worked for six months at the call center, alongside as many as 150 others. He said that after initial training, he went on to transcribe live messages.
“It was done 100% by people,” he claims, adding, “We heard the message from the very beginning to the very end. Love messages, secret messages, messages with sexual content, even people threatening to kill each other.”
Mr. Lucilius, however, seems to have worked for SpinVox under the auspices of RAYA. A Facebook account by that name has commented on the group page (this may be how the BBC reached him) and a “Kareem” is pictured sleeping or pretending to sleep in a photo. If he did work at RAYA, this raises some questions because SpinVox claims that company never saw any live messages.
On the other hand, a post on the Facebook page from April 2008 reads, “Mabrook for those who made it to the Live Session..You guys kick ass!” A second post in June 2008, from the same person, clearly identifying himself as being from RAYA, asks if the UK workers if their servers are down, implying that RAYA may have been processing live messages.
Whatever the situation, it is not helped by the fact that SpinVox’s new and rapidly expanding technology has some secrecy surrounding its center, leaving room for doubts, speculation, and innuendo.
The company won’t say, for instance, how much human intervention is involved in transcription, describing the actual proportion of messages automatically converted as “highly confidential and sensitive.” Rather, it says that it requires only a few hundred agents per market to convert messages without learning assistance. In Argentina, where SpinVox has 10 million customers, it has less than 70 call center staff, suggesting it’s not feasible for SpinVox to operate without some level of automated speech recognition, the Guardian suggests.
Ms. Lyons adds that SpinVox has “reduced human assisted learning to just 2% of what it was when we started.”
What that means without a clearer view of how many messages SpinVox processes with human assitance, however, leaves much to be answered. As my brother Adam B. would say, “This is a real brain buster!”
