Skip to main content

I spent this past summer working with Professor Daniel Swingley, the director of the Penn Infant Language Lab, performing computational analysis on Vowel Categorization in a Conversational Corpus. The goal of my project was to be able to use this corpus in an attempt to see if it would be possible for an algorithm to recognize the correct vowel categories. A given vowel, like an “e” or an “a”, sounds somewhat different each time it’s said by a speaker.  In fact, some “e”s are, phonetically, more similar to other vowels than they are to most “e”s.  This raises a learning problem.  If all vowels sounded the same, it would be easy to learn how many vowels a language has, and what each one sounds like.  But because the sounds are so variable, it is hard to tell what a language’s vowel categories really are.  Researchers have struggled to work out how babies might discover their language’s vowel categories. When you look at the average way a vowel sounds in a word, however, the hypothesis is that a computer would be able to recognize the correct vowel categories. An analysis of the Spanish language was successful so we applied similar logic to an adult English language dataset as it was recorded with higher quality than existing English infant language datasets. We used an automatic process to extract vowel measurements, but this process sometimes makes errors. In order to remedy those errors, we applied alternative analyses and regression techniques to repair as many errors as possible. When analyzing the data, we found that clustering over word averages was helpful in identifying vowel categories, yet performance varied based on speaker. The token analysis, as expected, had much more noise and did not yield as successful categorization. 

Through my research experience, first and foremostly, I was able to gain a better understanding of the research process. Before my PURM experience, I had very little knowledge of the way research was conducted and now that I have a better understanding, I can adjust my expectations for any future research endeavors. Not only was I able to understand the research process, but I was also able to gain knowledge in both linguistics and engineering. Learning about formants was something I don’t think I would have ever come across in any of my engineering courses and I’m glad to have learned about this fundamental of vowel analysis. In addition to non-engineering skills, I was able to develop skills in both R and Python – skills that improved me as a computer science major and will have great use for the future. 

I spent this past summer working with Professor Daniel Swingley, the director of the Penn Infant Language Lab, performing computational analysis on Vowel Categorization in a Conversational Corpus. The goal of my project was to be able to use this corpus in an attempt to see if it would be possible for an algorithm to recognize the correct vowel categories. A given vowel, like an “e” or an “a”, sounds somewhat different each time it’s said by a speaker.  In fact, some “e”s are, phonetically, more similar to other vowels than they are to most “e”s.  This raises a learning problem.  If all vowels sounded the same, it would be easy to learn how many vowels a language has, and what each one sounds like.  But because the sounds are so variable, it is hard to tell what a language’s vowel categories really are.  Researchers have struggled to work out how babies might discover their language’s vowel categories. When you look at the average way a vowel sounds in a word, however, the hypothesis is that a computer would be able to recognize the correct vowel categories. An analysis of the Spanish language was successful so we applied similar logic to an adult English language dataset as it was recorded with higher quality than existing English infant language datasets. We used an automatic process to extract vowel measurements, but this process sometimes makes errors. In order to remedy those errors, we applied alternative analyses and regression techniques to repair as many errors as possible. When analyzing the data, we found that clustering over word averages was helpful in identifying vowel categories, yet performance varied based on speaker. The token analysis, as expected, had much more noise and did not yield as successful categorization. 

Through my research experience, first and foremostly, I was able to gain a better understanding of the research process. Before my PURM experience, I had very little knowledge of the way research was conducted and now that I have a better understanding, I can adjust my expectations for any future research endeavors. Not only was I able to understand the research process, but I was also able to gain knowledge in both linguistics and engineering. Learning about formants was something I don’t think I would have ever come across in any of my engineering courses and I’m glad to have learned about this fundamental of vowel analysis. In addition to non-engineering skills, I was able to develop skills in both R and Python – skills that improved me as a computer science major and will have great use for the future.