Speech-to-Text Technology Cruder Than Advertised

Speech-to-text technology is nowhere near as good as many vendors claim. Often, operators are surreptitiously involved.

According to this BBC article, one of the speech-to-text conversion companies--Spinvox--has been found to convert voicemail messages using human operators instead of "advanced speech recognition software." Patent filings by Spinvox reveal the depth of human involvement. Spinvox's valuation has dropped 90% since this news has gotten out.

Plenty of such vendors do the same thing. Their contracts include fine print about being able to open messages for "quality control purposes." They don't state that in fact every message is indeed opened for quality control and editing.

The problem for the speech-to-text industry is that its automatic technology, while evolving rapidly, can't do near-perfect translation. The human involvment means that users of speech-to-text have to live with numerous typos and other misinterpretations.

In the case of vendors like Spinvox, a shell game had emerged whereby vendors were trying to get ahead of the pack by employing vast teams of "editors" in Pakistan and other countries. No one knows how long this will have to continue, until automated technology is good enough. Thus many speech-to-text companies have had to maintain these pools of human assistants--along with the associated cost burdens--for an indefinite time. As a result, some companies have been bleeding cash.

In a sense, employing human editors would seem to be an innovative and useful approach. But sometimes this simply leads to everyone playing the same game. All the spin quickly leads to an overinflated market and things crash.

Customers would be wise to read through the fine print of vendor claims if they are using speech-to-text services such as that of Spinvox. At the end of the day, if customers are happy with the service, and fine with the security risks of having people halfway around the globe reading your email, it’s not necessarily the end of the world to have your voicemails transcribed to text and sent to you in under a minute.

...David Sengupta

Post a comment

You must be logged in to post a comment. To comment, first join our community.