After several days of no new participants and my aversion to continuously spamming Facebook I have correlated the first set of data from the survey. Thank you everyone!
I won't post all the results here, but I will give you a look at the results for one of the more important questions (#9).
Suppose you are given a microphone and a program displaying the image above.
How would reword the command "move the red square up"?
Replies: There were 42 unique replies.
| square up move red up red square up move square up move the box up red square go up shift the box up square up 1 inch square move north move square to top move the square up push the square up | move the red box up move the red one up red square slide up shift red square up shift the square up slide red square up direct red square up move the red thing up move red square north red square move north push the red square up move the red object up | shove the red square up slide the square box up make the square go north move the red square north translate the red square up make the red square move up move the red square up 1 inch move the red square vertically move the circle below the square click on red square to move it up move the red square to location X relocate the red square towards the top |
| relocate the square in an upwards motion move the red square to the top right corner position the red square at the top of the screen move the red square to the right of the blue circle move the red square to be on the same plane as the blue circle in 5 equal increments move the red square straight up stopping at the black frame | ||
The power of suggestion is definitely a problem in collecting this data, and it is part of the reason the survey is so vague in the first few questions. By asking for variations of "move the square up" most of the results had 'square' in it instead of some other word. I'll have to crunch the numbers on all the data to see if this type of thing is common and/or related to part of speech.
The plan is to take this data and the data from other questions and use it to improve speech recognition. In a nutshell I intend to use the Stanford Grammar Parser to match the parts of speech of each word in each command in the list above to the parts of speech of each word in "move the red square up", the known command. Once I know which words are serving the same function grammatically in the command I will then use WordNet to see how close nouns match known nouns semantically, verbs w/ verbs, etc. This will give me a numeric score for how well the command 'spoken' and the known command(s) match. If the score is high enough the matching known command will be executed. I hope to create a generalized algorithm that works well without the need to collect large data sets like this.
This algorithm would augment a speech recognition system that only knows very specific known commands...essentially if this algorithm recognizes (on average) more commands accurately than inaccurately, and is able to reject unknowns, even a single success would be a huge improvement.
This algorithm would augment a speech recognition system that only knows very specific known commands...essentially if this algorithm recognizes (on average) more commands accurately than inaccurately, and is able to reject unknowns, even a single success would be a huge improvement.
I'm designing the software in such a way that in the next 'survey' participants will be able to actually use the system so they can see the commands work. I hope to also come up with a method of showing graphically how the score was arrived at but that might not make it into the next survey...
Thanks again everyone. Stay tuned!
Posts
