I just read a TechCrunch post on companies that should exist but dont, which reminded me of an idea I had years ago during the dawn of the internet age, to use Internet technologies to farm out translation and transcription work to low-wage areas, not just India and China but the rural United States, etc.

One of the companies mentioned in the post is a podcast transcription service. It occurs to me that the new Amazon web service, Amazon Mechanical Turk, could be the key to implementing just such a service.

The service has some shittly limitations that I hope will go away:

• The requestor fixes the price for the service. An obvious alternative pricing mechanism would be a reverse auction wherein the requestor specifies a maximum price (equivalent to the reserve price in a standard eBay auction), and workers bid on the work

• There is no sandbox, so playing with the service is somewhat complicated

• The requestor funds an account with the amount required to pay for the request, along with commission, before the request can be submitted. Upon receiving the response, the requestor can accept or reject the work product. So, if an intermediary service were acting as a broker (as a podcast transcription service would), the broker takes on some risk by fronting for the service, or the broker must require the client to similarly prepay.

• The qualification system, whereby requestors can test the qualifications of prospective workers, is very limited. For transcription purposes, I’d want to attach an MP3 file of a conversation to be transcribed, maybe a link to some instructions on how the transcription should be formatted, and a free text box where the transcription should go. Of these, only the free text box is supported now.

• There is no support for competition among workers. You can allow a task to be performed by multiple workers, but the expectation is that anyone who follows the instructions will receive payment. So, telling three people to transcribe a conversation means that all three are supposed to be paid, even if one is clearly better than the other two. This needs to change if AMT is to be at all viable.

• Apart from the qualification system, there’s very little ability to restrict tasks to specific workers. In the case of transcription, you’d want the requestor to have some subjective feedback measure, and you’d want that to help determine if the worker will get work from you again. Or, perhaps you would want to allow the requestors to decide if they want to pay for a highly regarded translator, or only a mediocre one.

  AMT intentionally hides the ‘man behind the curtain’, operating instead as a black box into which work orders go, and from which completed work emerges. This is an unfortunate decision; individual accountability, and the ability of an individual to excel above others, is very important in a competitive service economy.

• Even the qualification system is somewhat limited, in that workers must request a qualification before it can be granted. So, you can’t, say, create a TopTenPercentWorkers qual and grant it programmatically to the best 10% of your workers; it must be requested and adjudicated on a per-worker basis.

• Ideally, the requestor (or more accurately, software operating on behalf of the requestor) could review work at submission time, as a second-stage validation step, and provide feedback which the worker could incorporate into the work product. For example, when a worker inputs a transcription, it could check for adherence to transcription format conventions, case, spelling, etc.