VoxForge
VoxForge
I was looking at different ways to generate corpora for training acoustic models, and it occurred to me that using Mechanical Turk might be a good way to do it:
For those not familiar with Mechanical Turk, it is a system for distributing simple tasks (they call them Human Intelligence Tasks) among many workers, each of whom get paid a small amount per task.
For instance, one HIT could be submitting audio via the java applet at http://voxforge.org/home/read. Another task could be to transcribe the resulting audio.
So, you could pay people 20 cents to read 10 sentences, and 20 cents to transcribe 10 sentences (to double check the work). This would come out to around 40 cents per minute of transcribed audio, or 24 bucks per hour, which is pretty cheap. Looking at the current set of tasks there, I think those prices are reasonable.
That way, instead of having to sit there and transcribe audio for an hour, people could donate $25 bucks over paypal and get the same result.
--- (Edited on 12/5/2008 5:57 pm [GMT-0600] by iragsdale) ---
Hi iragsdale,
>using Mechanical Turk might be a good way to do it
I heard of some speech recognition companies using something similar to collect speech... I guess the donation collection could be set up on the VoxForge SourceForge page (using the SourceForge "Donate: feature - which is not currently set up...).
However, I am not sure that enough people are willing to donate money (to pay for others to record their speech) to make it worthwhile to set up... If people are interested, let me know in a reply to this post...
thanks,
Ken
BTW we have collected about 56 hours of speech, so that would have cost us about $1,344 (based on your mturk estimates).
--- (Edited on 12/8/2008 11:35 am [GMT-0500] by kmaclean) ---
I setup a mturk account. I threw 100$ on and well see how much I can get for as cheap as humanly possible.
Oddly enough though, I think theres a mturk bug. I believe your not supposed to be a requester(aka those who pay to get tasks done) unless your an American. Yet my account was able to be funded and when I go to personal information it appears blank except the country says USA which I never put. I think what happnd is that I registered as a worker(foreigners are allowed) and then went to the request section.
I am going to send amazon a email, if its ok I will put up the hit(what they call a task) asap. Worst case I can paypal or use amazon payments and transfer over that 100 for one of you americans to do it.
--- (Edited on 3/7/2009 12:17 am [GMT-0600] by ecret) ---
Hi Ecret,
>I setup a mturk account. I threw 100$ on and well see how much I can get
>for as cheap as humanly possible.
Very cool! Let us know how you make out.
thanks,
Ken
--- (Edited on 3/7/2009 1:27 pm [GMT-0500] by kmaclean) ---