Phoenetic Variation in the New York City Community
Advisor: Caitlin Collins
What defines a New York accent? One common element is rhoticity: the presence or lack of the “r” sound at the ends of words, like in the pronunciation “New Yawk”. However, some people have stronger or weaker accents; for example, older people usually have stronger ones. This project will address the lack of research about if the strength of New York accent, measured by the amount of rhoticity, correlates with someone’s demographics, like age, ethnicity, and gender. I will study speech data from the CUNY CoNYCE (Corpus of New York City English) to quantitatively measure and compare accents among speakers. This will be used to determine and measure how much New York speech is changing over time and how it’s influenced in different groups.
Project Blog
Week 0: Introductions
Hello world! My name is Maxwell Hyun and I’ve been at Basis Independent Brooklyn since 2nd grade. For my Senior Project, I will be studying phonetic variation within the New York accent. With this project, I hope to supplement the lack of existing research on the topic, particularly across NYC as a whole as opposed to within each individual borough. Additionally, this research will help document the traditional New York accent, which is becoming less and less common over time.
I’ve been into linguistics for years, watching videos on YouTube or reading Speculative Grammarian (the foremost publisher of scholarly research in the lesser known and in my opinion severely neglected field of satirical linguistics), and finishing an introductory Coursera course. I enjoy learning other languages, including quite a bit of Chinese, a good amount of Japanese, some Korean, and a little ASL. I also like to learn about conlangs, short for “constructed languages”, which are languages made by people rather than naturally arising. They’re very interesting to me because each conlang can display interesting new ideas about language, like one in which you have to express everything as combinations of relatively few simple words (jan pi toki pona li lon ala lon?). I think this project will be a fun opportunity to learn more about linguistics. I also enjoy coding, and have incorporated that into my project as well.
Specifically, I’ll be measuring the New York accent through coda rhoticity, or the presence of the “r” sound at the ends of syllables, such as in the words “store” or “part”. I’m going to work with my on-site advisor Dr. Michael Newman, the Chair of the Department of Linguistics and Communication Disorders at Queens College. He will help me write programs to analyze data from the CUNY CoNYCE (Corpus of New York City English), a corpus with recordings from 200+ interviews with different people across NYC. Then, we’ll analyze this data to see if/how it correlates with people’s demographics, including age, ethnicity, and gender. For my final product, I will be writing a research paper on my findings.
If you’re curious and/or daring enough, read more about my project in my proposal.
Week 1: Initial Research on Phonetic Analysis (or, What the heck is a formant?)
Hello, all! This week marked the official beginning of my project and on-site placement. Things are off to a good start—Dr. Newman has been helpful in planning out the path ahead and meetings have been productive. So far, I’ve done some research into my course texts to nail down my methodology; I read Villareal’s “Gender separation and the speech community: Rhoticity in early 20th century Southland New Zealand English” from Language Variation and Change, and Turton and Lennon’s “An acoustic analysis of rhoticity in Lancashire, England” from the Journal of Phonetics. Each used a different method to measure rhoticity: Villareal used machine learning to categorize sounds while Turton and Lennon used programs to measure formants. Turton and Lennon’s paper will serve as a good model for my own research. Ultimately, Villareal’s turned out not to be as relevant to my project since its method preassumes that rhoticity is binary—I don’t. Even so, it’s still useful for the ideas it presents if not for methodology.
By the way, for the majority of people who don’t know what I’m talking about when I say “formant”, a vowel sound is essentially made up of a few different pitches put together, where each pitch is called a formant. Different positions of the tongue in the mouth change how high each formant is, so the sound of a vowel is defined by the pitch of each formant and how they relate to each other. For example, here’s a graph of the vowel in “thirty”.
Each colored line shows the path of one formant. The blue line on top moves down over time, showing the change from “i” to “r”.
This is a good segue to what else I’ve been doing, learning about the programs I’ll be using to analyze these formants. Right now, these are Praat and FastTrack. Praat is a commonly used program used for speech analysis in phonetics, and FastTrack is used to track formants accurately. In fact, I used these two to generate that formant graph! This week I also began some online courses about research ethics on CITI’s (Collaborative Institutional Training Initiative) website. These courses are required for me to access the CoNYCE’s data.
Looking forward, I’m planning to learn more about FastTrack and continue the research ethics courses. Once I complete them and get access to the CoNYCE sound files, I can look into processing them. I’ll also be reading Lipari’s “The Acoustics of Borrowed /ɚ/ in Quebec French” to look more into phonetics research.
Week 2: Preparations for Analysis (or, Praat Scripts and Manuscripts)
Hello, all! This week, I’ve made progress preparing to analyze the data. First, I’ve been looking into specifics on using Praat and FastTrack, and have found success with each. Praat, it turns out, is actually very easy to use. It can perform many very useful operations, and it didn’t take much digging to find, for example, how one might import a sound and measure the minimum F3 (3rd formant—see Week 1). The Praat scripting language is also very intuitive. It’s generally structured such that if you can do something in the GUI, then you’ll be able to automate it with a script without much trouble. With FastTrack, I’ve conducted some tests going through the methods of formant measurement I’ll be using, and managed to get data out of some test audio. Now, the challenge from there will just be to automate the process for multiple audio files at a time.
I also read through Lipari’s “The Acoustics of Borrowed /ɚ/ in Quebec French”, and it turned out to be very interesting. For space, I’ll refrain from describing the entire thing in detail, but it discusses the rhoticity of English loan words with “r” in French—some people replace the sound with a French “r”, while others keep it. There are some elements of its methodology, like formant normalization, that I could apply to my own study.
In other news, I finished the CITI courses this week, so now I can officially say I know how to conduct research ethically without mistreating the human subjects of such research. With that, I have gained full access to CoNYCE’s files. These include recordings of all the interviews and a giant spreadsheet of demographic information about the interviewees, which had so much data—age, gender, race, ethnicity, place of birth, education level, home language, etc. of every single person—that when I first saw it I was momentarily laughing to myself like a mad scientist thinking about what I could do with it all. While some of the interviews have transcriptions already made, the majority don’t, so my next step is actually to produce transcriptions for the rest.
Finally, Dr. Newman has introduced me to another paper to read, Love & Walker’s “Football versus football: Effect of topic on /r/ realization in American and English sports fans”, so you can anticipate hearing about it next week.
Week 3: Producing Phonetic Transcriptions (or, “Why isn’t MFA working?!”)
Hello, all! This week has been interesting. The main task accomplished was making transcriptions for the remaining CoNYCE interviews. First, I’ll need to explain a little more about what I specifically mean by that.
To analyze recorded speech, I need to have a phonetic transcription. Unlike a regular transcription, a phonetic one records every phoneme, or speech sound, of each word. For example, for the sentence “I was born in New York”, the phonetic transcription might look like /aj wəz bɒɹn ɪn ɲʉː jɒɹk/. This is written using the International Phonetic Alphabet, which has one symbol for each possible phoneme, unlike the English alphabet where one letter could sound different depending on context (e.g. “c” in “cat” vs. “cite”; these would be /kat/ and /sajt/ respectively—unambiguous). Also, my transcriptions need to be time-aligned, which means they record the start and end times for each sound. This is necessary because it allows me to extract the audio of each instance of the /r/ phoneme for acoustic analysis.
Back to my project, while all of the CoNYCE interviews already had normal transcriptions, only some had the necessary phonetic, time-aligned transcriptions, so I had to make these for the rest. Of course, I wouldn’t be transcribing a hundred interviews by hand, rather there are programs to do it for me. I chose to use one called the Montreal Forced Aligner, which takes the normal transcription and audio as input and produces a time-aligned phonetic transcription.
However, getting MFA to work became my first big challenge. While testing it out, I found that for some interview files it worked fine, but for others it failed with a mysterious error. After days of testing and a meeting with one of Dr. Newman’s colleagues who is more familiar with MFA, I found that it actually worked normally on other computers, just not my laptop. This seemed to suggest that something about my computer’s environment was conflicting with MFA. So, I installed it on a separate Linux environment on my computer, and that fixed the problem.
Now, as I mentioned last week, I will discuss the paper I read, Love & Walker’s “Football versus football: Effect of topic on /r/ realization in American and English sports fans.” Essentially, they found that speakers of British and American English would speak with more of an American accent (measured by high rhoticity) when talking about American football, and more of a British accent (low rhoticity) when talking about English football/soccer. I personally find this very fascinating, as I didn’t think that accents were so flexible and variable in the moment. It also relates to my project, as it is another example of measuring accent variation via rhoticity.
Now that I have the transcriptions for all the remaining interviews, my next step is to use FastTrack on them to get formant data, and then begin analysis of this data.
Week 4: Aggregating the Data (or, Please Check for Typos Before Publishing)
Hello, all! This week, I finished the final step to get usable data out of the interview recordings. First, I ran FastTrack on the audio and transcription files. I had to process it in 3 separate batches, for two reasons: First, the transcriptions we already had and the ones I made are written using different writing systems; the former uses the IPA (e.g. /ðɪs ɪz ði ɪntərnæʃənəl fənɛtɪk ælfəbɛt/), while the latter uses ARPAbet (e.g. DHIH1S IH1Z IH1N AA1RPAX0BEH2T). Second, I’m measuring two types of /r/s, ones which act like vowels, like in “nurse” where the entire vowel sounds like /r/, and ones which act like consonants, like in “ear” where the syllable starts with the “e” vowel and then ends with an /r/. These two have to be measured individually since they are sometimes realized differently, especially in the New York accent. A common example is Bugs Bunny’s accent, where for the "vowel /r/", he'll say words like “turn” as “toin” or “learn” as “loin", whereas he'll just drop the "consonant /r/". Also, in ARPAbet, these vowel /r/s have their syllable stress marked, but the consonant /r/s don’t. Therefore, In the end I had to do one batch which looked for vowel “ER”s marked with stress in ARPAbet, one which looked for consonant “R”s without stress in ARPAbet, and one batch which measured all “r” vowels in the IPA (also without stress).
Doing this got me a handful of CSV (Comma-Separated Values) files from each batch, which I then had to combine. To do this, I wrote a python script which goes through each CSV from each batch and organizes it all into one output CSV, along with demographic data from another source. In the process of doing this, I also found that I needed to do some data cleaning—there were a few mistakes, like misspelled names or other information that needed to be corrected for everything to work properly. Once I did this (as well as, of course, copious debugging to get my script to return properly constructed and formatted output), I finally got the final output file with all of the data aggregated! Now, I just need to analyze this data statistically to get my results.
Week 5: Performing Statistical Analysis (or, RESULTS!!!)
Hello all! This week, I got some real results! I started with learning the relevant statistical analyses and how to do them using the statistical software Jamovi. First, I imported the data and calculated the difference between the third and second formants (F3-F2), a typical measure of rhoticity, for each /r/ instance, or token. A lower difference corresponds to greater rhoticity and vice versa. I also categorized the tokens into four types: “Nurse”, “Shwar”, “Coda”, and “Linking”. Each type represents an /r/ token from a different context. “Nurse” and “Shwar” tokens are what I referred to last week as “vowel /r/s”; “Nurse” tokens are stressed, like in the word “nurse”, and “Shwar” tokens are unstressed, like in the word “after”. “Coda” and “Linking” are both “consonant /r/s”, but “Linking” tokens are a special subset. A linking /r/ is one which comes at the end of a word where the next word starts with a vowel, like in the phrases “car owner” and “four apples”. Linking /r/s are still rhotic in some accents which would otherwise drop the /r/, hence why they’re measured separately.
Then, I plotted token density by F3-F2, to tell whether rhoticity is binary or not. If it is binary, the expected plot would be bimodal, with one peak for rhotic tokens and one peak for non-rhotic tokens. However, the plot I got was unimodal, giving me my first result by showing that rhoticity is, in fact, non-binary. In other words, unlike what previous studies assume, the distinction between a present or dropped /r/ is arbitrary, and there’s just a continuum of “/r/ presentness”.
After that, I used a linear mixed effects model to look for relationships between demographics and rhoticity. This is like a linear regression, but allows for multiple factors and groupings of the data, which is what I need to analyse all of the different demographics together. In the end, I found the following: First, rhoticity decreases by age. This lines up with previous findings and expectations, as there is a known trend away from the traditional /r/-dropping New York accent in younger people. Second, women’s speech was more rhotic than men’s. This makes sense because it has been found that in the case of a change over time like this, the men will lag behind. Third, the different /r/ types were pronounced differently: “Nurse” tokens were much more rhotic, and “Linking” tokens were somewhat more rhotic, than “Coda” and “Schwar”, which were the least rhotic. Finally, race actually didn’t have an overall effect on rhoticity, which I was surprised to see. However, different race groups did differ in how they treat the /r/ types. White and Latinx people showed a strong linking /r/ and “Nurse” rhoticity, Black and mixed race people had a rhotic “Nurse” but no linking /r/, and East Asian and South Asian people had no significant difference between any of the /r/ types.
My next step is to continue analysis, including comparing the rhoticity distribution of individual speakers to see if individual rhoticity affects it.
Week 6: Further Statistical Analysis (or, “The Fifth Element”)
Hello all! This week, I continued with statistical analysis. As I mentioned in my last post, I performed separate analyses on only speakers of especially high or low rhoticity to see if it had any effect on the rhoticity distribution; turns out it didn’t. I also tried plotting different combinations of variables, and found another interesting outcome—when comparing the effects of gender and age on rhoticity, I found that men tend to speak as if they were about 40 years older compared to women. To put it another way, an average 20-year-old man would speak with a similar accent to the average 60-year-old woman.
The most interesting thing, however, was when I added a new type of token. Last week, I described the four types of “r”s I had been working with: “Nurse”, “Shwar”, “Coda”, and “Linking”. All of these occur at the ends of syllables. However, FastTrack actually measures all instances of “r”, including at the beginnings of syllables; these are called “Onset” tokens. However, as no variation would be expected between these “Onset” tokens (people seem to pronounce the “r” in words like “run” about the same way regardless of accent), I was filtering them out of previous analyses. This time, I decided to include the “Onset” tokens in an analysis just to see what would happen, and found something very strange. They seemed to vary right along with the other types of “r”.
This doesn’t make very much sense intuitively, so my first thought was some kind of mistake in measurement. Maybe some tokens are being miscategorized, or there’s some other error. Maybe there really is variation in “Onset” tokens as well, and nobody noticed? Either way, I’ll be trying to figure that out.
Week 7: Collection of Additional Data (or, Please ORGANIZE Before Publishing)
Hello all! First, I’ll address the mysterious results with the “Onset” tokens. After discussing with Dr. Newman, we agreed it seems the results aren’t spurious because the values do make sense. If it were an error in measurement, the “Onset” tokens might be much more rhotic or much less rhotic than the other tokens. In reality, they were about the same as “Nurse” tokens, which aren’t usually weakened, so there’s probably no mistake. This really does imply that basically all /r/ tokens, even at the beginning of words, are weakening over time. This is very counterintuitive to say the least, as most people assume that these /r/s never vary.
About the rest of the work I’ve been doing, it has mainly been collecting more data. The corpus files are somewhat…disorganized, so my initial dataset was actually missing many interviews. This week and last week, Dr. Newman and I devoted much time to combing through the many different folders and lists of people, downloading more recordings and transcripts, renaming files, fixing typos, reprocessing data, and aggregating results. In the end, the results didn’t actually change much with the new data, but it did reduce the error margins. This is important to make our conclusions more precise and accurate.
My next task is to collect all of my results and conclusions and compare them to those of my model studies. Stay tuned for next week’s entry!
Week 8: More Processing Necessary (or, “Always Check Your Assumptions”)
Hello all! This week, Dr. Newman and I had an important realization about my results. We were looking at how different “r” types compared to one another, and saw that “Nurse” tokens were overall more rhotic than “Onset” tokens. This doesn’t make much sense—as I mentioned before, it is usually assumed that “Onset” tokens should be the most rhotic and never weaken, because people don’t typically observe dropped “r”s at the start of words. We had noticed this contradiction already, and chalked it up to an issue with measuring the “Onset” tokens.
However, this actually seems to be part of a deeper problem with our analysis. This entire time that we’ve been comparing the different types of token, we have been making the underlying assumption that the F3-F2 measurement means the same thing for all types. But really, it doesn’t correctly represent the amount that the “r” was weakened. For example, if a “Shwar” token was fully weakened, as in a British accent, it would become the vowel /ə/—found in General American English at the ends of “arena” and “Laura”—while if a “Nurse” token was fully weakened, it would become /ɜ/ (not otherwise found in General American). Because the different “r” types weaken to become different vowels, comparing them directly with each other is inaccurate. Instead, each should be compared with its corresponding vowel to determine how weakened it has become.
To this end, I now have to go back through all of the recordings and measure F3-F2, our main metric of rhoticity, for all of the normal vowels like /ə/ and /ɜ/. This way, if some vowels sound more rhotic, we can correct for that in our analysis of “r”. On the bright side, this will also be very useful data to have for future study by myself or others. On the dark side, running FastTrack on the hundreds of recordings will take a very, very long time. Hopefully I can mitigate this by having it run in the background as much as possible. Either way, we’ll see how far it gets by next week.
Week 9: Continued Processing (or, “So Much Waiting…”)
Hello, all! Last week, I mentioned I’d be running FastTrack on the corpus to measure the normal vowels. I’ve been doing this in batches, to make sure that if anything went wrong in the middle and I had to restart, I wouldn’t lose too much progress by having to rerun it. This turned out to be useful when, as I predicted, something went wrong in one of the batches and I had to restart.
It seemed that FastTrack was failing to process one of the files, although I don’t quite understand the error produced yet. I took that file out and reran the rest of the batch, and I plan to discuss this at my next meeting with Dr. Newman.
Overall, the process is going very slowly, so I might need to change my plans in order to finish in time to use the findings for my project. I could switch to just measuring the pertinent vowels for my work with rhoticity, and leave out the rest of the vowels that I don’t need, or only process some of the files, just gathering enough data to establish the vowel baselines I’m looking for. Either way, I’ll have to consider how long the new option would take and how long I expect the rest of the work to take, compared to the time I have. Once again, I’ll talk about it with Dr. Newman and make a plan by next week.
Week 10: Final Processing (or, “Team Work Makes the Dream Work”)
Hello, all! This week, I made plans for everything going forward. In last week’s post, I mentioned that I had to figure out how to finish running Fast Track on everything in time; this week, I met some people who can help.
Me and Dr. Newman aren’t the only ones using the CoNYCE, but until this week we hadn’t all met each other. In total, there are four others: one college student, who has been working with Dr. Newman for some years already on a separate project studying vowels, and three high schoolers like me, who just joined recently to learn more about linguistics and help out however they can.
The plan we came up with is to distribute the remaining files that need to be processed among the high schoolers, and then have them run Fast Track on their own computers. Once everyone finishes processing their sections, I can combine all of the outputs. Splitting up the work like this will hopefully allow us to go four times as fast, and get the last 150 files done by the end of this week. Once this is done, not only will I have more accurate data to use in my project for measuring “r”s, but this data will also be available for other people to do their own research with.
In addition to all this, the official end of Senior Projects is approaching, so over the next week-and-a-half I’ll be working on my final presentation and my deliverable, which is a research paper documenting my work.
Week 12: Conclusion
Well, this is it: my last post. This week, I submitted my paper and gave my presentation; overall, I think they both came out pretty well. That means that this post also marks the end of my Senior Project as a whole. It’s a bittersweet moment: I’m glad to be free for the summer, but I have been enjoying working on this project for the past three months.
Although it’s been a lot of work, with many challenges and setbacks along the way, it’s also been a chance to choose my own fate, so to speak. I got to study a topic that I enjoy, so working on it never felt like a chore. I got to decide how to spend my time every day (at least, outside of my weekly track practices), so my schedule never became too overwhelming. Because of this, hopefully I’m at least a little more prepared for the independence of college life now.
Speaking of preparing for college, I think I’ve definitely learned a lot from my project that I’ll be able to use in college, like linguistics concepts and tools as well as how to do original research. I also have new connections which could come in handy.
While this may be the end of my Senior Project, I’m still not done with this project. I’m going to keep working with Dr. Newman and submit my work for consideration to present it in a conference next fall. Not to mention there’s always more analysis to do on the data.
Well, thank you for reading and following along on this journey with me. Goodbye all!