DREU SUMMER 2012

Sarah Ita Levitan

HOME ABOUT ME MENTOR PROJECT JOURNAL REPORT

Journal

Week 8
This week was exciting! I finished retrieving the results for the balanced set of HITs, and got some statistics about the results (how many were noisy/clean, which sessions they were from, were they mostly from the beginning, middle, or end of the session etc.). We found that the middles were the cleanest segments, but there were not particularly significant results. We decided to take a new approach, and load complete, consecutive sessions to AMT instead of randomized segments, in order to make the tasks more interesting and meaningful. I loaded a complete session to AMT, and while waiting for the HITs to be completed, I wrote a script to sort the results and filter out desired fields from the messy results file.
I then wrote another script to take the new and improved results file, and mark the original csv (comma separated values) durations file (lists each ipu - session number, segment number, and duration) with a new field - noisy or clean. I also wrote a script that takes the marked csv file and uses it to create a text grid (file used by praat) where the segments are labeled as noisy/clean. This will be useful when running praat scripts on the data to extract features of speech - we will simply skip over the ipus that are labelled noisy.
AMT was going pretty slowly, so we recruited some fellow lab members to do a couple of HITs. We also brainstormed about how to attract more turkers to do our HITs. Someone suggested that we remove the qualification exam, and after I tried this, the final 10 HITs were completed within minutes! Instead of relying on the qualification exam, we now have to spot check the results to make sure they are being done properly. We are also adding some "gold standard" questions (which we know if they are noisy or clean) and will automatically reject any turkers that get these questions wrong.

[PREVIOUS] [NEXT]