One of the things I saw at ETech was castingWords.com - a company providing transcriptions of podcasts. Their service sits on top of Amazon's Mechanical Turk service, which is a web service for tasks that require human intervention. The Mechanical Turk is designed for tasks like "Select the best photo of the shop from this set" - which are easy for a human but difficult for a computer. There's a way of submitting tasks and setting the payment and there's a way to see tasks that have been set and selecting ones to complete - any user can then decide to do a task and get paid if successful. There's even an API for including these manual tasks in your (asynchronous) code.
Part of castingWords business is that they include the transcribed podcasts in a podcast search engine which you can see on their homepage.
So what happens with castingWords is that you give them a URL for a podcast and then they farm this out via the Amazon Mechanical Turk service to approved transcribers, who transcribe it and create plain text, RTF and HTML versions. Their current rate is $0.42 per minute which works out at approximately $25 or £15 per hour. Or, by my calculations, £360 per radio network per day.
So I submitted a BBC radio podcast, a Radio 4 Front Row interview with Steve Reich. It's just 10 minutes long so that's $4.20 in total and it took about 3 days to be done. Interestingly they give you a personalised RSS feed which keeps you informed as to their progress.
Here's the first bit of the transcribed interview...
Mark: Thanks you very much James, Trans-America's 15 acres around the country on Friday the shape of that film was a journey from New York to California by car, the same route taken often by train after his parents divorced. In spite one of the made best pieces by the American composer Steve Reich. Different trains come faster on those Amtrak trips taken by young Jewish Americans with the fatal rail journeys to Nazi death camps in Europe. Reich is 70 this year and has consistently worked with the media, employing audio playing at different speeds. He calls is phasing video tapes and often collaborating with choreographers. A new ballet with Akram Khan, variations for vibes, pianos, and strings, opens this week in London. During rehearsals I suggested to Steve Reich that the many concepts to honor his 70th birthday invited him to listen to his old work, How does he react to it ?
Steve: Well, by and large I love it the good news is that the old pieces sound great. I have my favorites and my less favorites. What is really gratifying is to go to best locations and instead of having my ensemble, to go and hear ensembles I have never met before or people I have never met before, people who might not have even been born when I wrote a lot of the pieces they are playing. And seeing that.. not only are they playing them, but they are playing them obviously with conviction, and with a great deal of mastery. And there is nothing, I think, that any composer would love more than to see that their music is played, is played well, and is played by musicians who love it, I mean - what more could you ask for?
Mark: And a question which in this country we call a desert island disk question I suppose is: Is there a particular piece that for you has worked out especially well?
Steve: Well I wouldn't single out one, but I would single out a sort of a greatest hits
[laughter]
Overall it's not bad; there are a few transcription errors (marked with bold above) and some domain-specific errors, i.e. something about "cultural, Contra Pudro" rather than "contrapuntal" or "Buck" rather than "Bach", but that said "Bela Bartok" and "
Tehillim" amongst others were correctly transcribed. You can in principle include some background and context information in your submission but I wanted to see how it went without my intervention. What's missing from my point of view is some timing information about what was said when. I've had an email discussion with castingWords and they do collect some timing information, you can see it in their
search results, but they don't think it's quite ready for inclusion yet. I also found this
comparison between a conventional transcription service and castingWords.
The availability of this kind of data suggests some interesting interfaces that we could build for radio programmes on the web providing alternative navigation and search. More to come soon.
Posts
