60 likes | 238 Views
Vowel.arff. Speech recognition Koen en Hraban. Problem!. How do we make a computer recognize vowels in speech?. Let’s solve…. We record fifteen different people pronouncing 10 different vowels each. We note features, like gender and more technical things.
E N D
Vowel.arff Speech recognition Koen en Hraban
Problem! • How do we make a computer recognize vowels in speech?
Let’s solve… • We record fifteen different people pronouncing 10 different vowels each. • We note features, like gender and more technical things. • After that, we store all the data in our extremely intriguing vowel.arff • We call in our little bird-friend WEKA, and let him do all the work for us…
We try: • Our decision stump didn’t really succeed quite well, so to speak… 11% accuracy. • How about a J48 then? We use 66% of the data to learn, and the rest to test. Nice: 71% accuracy. • BUT! We choose the best solution of course, as we are pioneers of the new knowledge!! It is:
CROSS-VALIDATION!!! • Now listen up, for the two people presenting this sheet will explain all the basics of this fine new genuine invention!!!
Results: • The decision stump was, of course, not the best way of predicting vowels. Though this was quite obvious from the beginning off. • The decision table, which we tried out in the mean time, did relatively well (68%) but far not enough for our research. • In the end we tried a few J48 trees: • Percentage split, 66%: it was 71% accurate. Nice, just not enough. • Cross validation? 10 folds: 81%!! That is very nice, don’t you agree? 4 out of 5 vowels are recognized correctly, all through ones and zeros… ain’t the technology just wonderful?!