# **Transcription, coding, and statistical analysis** These data are the product of personal interaction with members of the Puerto Rican community in Philadelphia, who generously gave their time and who willingly allowed an outsider into their home. The conversational data come from narratives of personal experience from members of the community, which were recorded in an environment comfortable to them. ## Sociodemographic information * **Survey (through Qualtrics™)**: This survey was designed to collect information regarding their age, sex, occupation, education, language background, and length of time lived in Philadelphia. * **Initial Sample**: 42 participants * **Exclusions** Six participants were excluded from the final dataset * lack of proficiency in English (n = 1) * not having lived in Philadelphia for the majority of their lives (n = 5). * **Final count**: 37 participants. Participants ranged in age from 19 to 63, with a mean of 36.54 years and a standard deviation of 13.24 years. |**Sex**| **≤ 1985**| **> 1985**| **Total**| |---|:---:|:---:|:---:| |Female| 18| 8| 26| |Male| 5| 6| 11| |**Total**| 23| 14| 37 ## Place of Residence The participants in the current dataset come from North and Northeast Philadelphia, and they can be subdivided into roughly three groups: 1. The southernmost cluster of participants, slightly north of Lehigh Avenue, lives in an area of long-standing Puerto Rican neighborhoods. 2. Directly north of the first group, along the so-called Broad Street corridor (e.g., Pew Charitable Trusts, 2011), extending northward to Olney, a suburb of North Philadelphia north of Roosevelt Boulevard (or, simply, “The Boulevard.”). 3. Another group of participants stretches northward and eastward, through Harrowgate, Juniata, and Frankford. [Participant Locations (Appx.)][1] These participants live in areas that, while predominately Hispanic, have also been influenced by recent patterns of migration within the city (Pew Charitable Trusts, 2011; US Census, 2010). In general, Hispanics are moving northward and eastward as whites move into areas closer to downtown Philadelphia. Each of the two branches of participants borders on areas with predominately African American residents, while the central group just north of Lehigh Avenue resides in a diverse space with whites, blacks, and Asians. ## Operationalizing socioeconomic status by occupation group Occupations were divided based on the Standard Occupational Classification Policy for 2010 from the US Bureau of Labor Statistics and previous research (Torres Cacoullos & Berry, 2018). Occupation groups consisted of Production, Service, Helpers/Aides, and Professionals. Additionally, Student and Unemployed were added as categories to account for individuals relying on government assistance and those whose careers were yet to be determined. * Data from aides and students were not included in the analyses by socioeconomic status. * Home health aides do not require specialized training, so in a sense they are unskilled service workers, but their clients—with whom they spend every working hour—may represent varied socioeconomic statuses. * Students are still to some degree financially reliant on their parents. **Speaker Sex by Occupation Group** |Sex| Unemployed| Production| Service| Professional| Aide |Student |---|:---:|:---:|:---:|:---:|:---:|:---:| |Female| 9| 2| 7| 3| 2| 3| |Male |4 |2 |3 |2 |0 |0| |**Total**| 13| 4| 10| 5| 2| 3| ## The speech data ### Conversational recordings: Narratives of personal experience #### Method: Sociolinguistic Interview All conversational recordings were obtained via traditional sociolinguistic interviews (Labov, 1981, 1984). * **Equipment**: Neck-worn cardioid microphones (Beyerdynamic [TG-H54C][2]) * **Interview Structure**: Based on [Q-GEN-II][3] 1. Demographic info: Age, schooling 2. Intimate topics: Crime, fears, dreams 3. Opinions on community: Public transportation, school system 4. Language: Impressions regarding different dialects in Philadelphia, linguistic attitudes In this interview style, the interviewer encourages the interviewee to speak, navigating the conversation according to the interviewee’s interests. In doing so, they focus less on the fact that there is a microphone nearby and more on the details of the narratives themselves, increasing the likelihood that their recorded speech represents their **vernacular**—the variety spoken when in a familiar, informal setting and when unmonitored. ### Read speech: Elicitation via [OpenSesame][4] * Participants read a series of words on a computer screen, presented via OpenSesame * 50 Stimuli for each of the three sound changes in-progress * 25 in favorable contexts for the novel variant * 25 were in other contexts. * For /ɔ/: * 25 stimuli had /ɔ/ as the stressed vowel and * 25 had /ɑ/, which was used as a reference vowel * 50 fillers with other vowels * Controlled for orthographic length, phonological length, lexical frequency (per million), orthographic neighborhood density (OND), and phonological neighborhood density (PND) |Variable Context|Ortho Length| Phono Length|Freq. per million|OND|PND|N| |---|:---:|:---:|:---:|:---:|:---:|:---:| |**EY-Raising**| Favorable|6.48|5.16|8.34|5.92|6.52|25| |*SD*| *1.42*| *1.31*| *1.43*| *4.57*| *5.51*|| |Other| 6.8| 5.72| 8.76| 7.16| 6.40| 25| |*SD*| *1.66*| *1.57*| *1.53*| *5.81*| *5.92*|| |**Canadian Raising**| |Favorable| 5.96| 4.4| 8.07| 5.16| 14.92| 25| |*SD*|*1.54*| *1.08*| *2.06*| *3.57*| *12.51*|| |Other| 5.48| 4.4| 8.18| 7.04| 14.8| 25| |*SD*|*1.26*| *1.19*| *2.06*| *3.76*| *12.36*| |**CAUGHT-COT**| |/ɔ/| 5.6| 4.28| 8.47| 4.08| 9.00| 25| |*SD*| *1.29*| *0.94*| *6.08*| *3.74*| *6.71*|| |/ɑ/| 5.36| 4.56| 8.82| 3.92| 9.04| 25| |*SD*| *1.32*| *0.77*| *4.99*| *4.07*| *5.15*|| |**Filler** | |Filler| 7.18| 5.82| 7.04| 2.38| 4.42| 50| |*SD*| *1.57*| *1.84*| *0.59*| *1.93*| *4.48*|| ## Transcription and coding ### Conversational speech * Transcribed in ELAN, beginning approximately halfway through the recordings at the start of a new topic. * ~10-12 min of speech were transcribed, stopping when a topic naturally ended. * The transcription process consisted of three phases. 1. Separate speech into Intonation Units (IUs) 2. Annotate each interval following the discourse-based transcription guidelines outlined in Du Bois et al. (1993) and utilized in the [New Mexico Spanish-English Bilingual (NMSEB) Corpus][5]. 3. Check transcriptions for accuracy, making manual adjustments/correctionswhen necessary. ### Read speech * Recorded word-by-word * Accuracy coded using PRAAT scripts and audiovisual examination. Only stimuli where the target word was correctly spoken were retained. ## Isolating tokens of the vowel targets ### Forced Alignment and Vowel Extraction (FAVE) [FAVE][6] (Forced Alignment and Vowel Extraction; Rosenfelder et al., 2014) is a program suite composed of two complementary components. 1. **FAVE-align** segments an acoustic signal into words and phonemes on the basis of a time-stamped transcription (forced alignment). 2. **FAVE-extract** takes that segmentation and extracts raw and normalized formant values for all vowels the model identified (vowel extraction). FAVE’s alignment algorithm is based on the Penn Forced Aligner (Yuan & Liberman, 2008), but it has been modified to better handle conversational speech with background noise and multiple talkers. Words and phones are output in [ARPABET][7] format. [Sample FAVE Output][8] ### Coding context and integrating sociodemographic information 1. Data were subset in R to include only stressed tokens of the vowels of interest (/eɪ/, /aɪ/, /ɔ/, /ɑ/). 2. A list of unique words containing those stimuli was exported and hand-coded by vowel and phonetic context 3. The coded list was then reimported and merged with the data. These data were then merged with relevant sociodemographic data (e.g., Year of Birth, Sex, Occupation Group) by speaker to arrive at the final dataset. |||Ey-Raising||Canadian Raising||/ɔ/-Lowering||| |---|---|:---:|:---:|:---:|:---:|:---:|:---:|---| |Speech Style|Sex|F: /eɪC/|D: /eɪσ/|F: /aɪ0/|D: /aɪV/|/ɔ/|/ɑ/| **Total**| |**Conversation**| Male| 242| 409| 570| 1,139| 467| 706| 3,533| || Female| 1,089| 1,807| 2,657| 5,871| 1,824| 3,341| 16,589| |**Read**| Male| 466| 465| 482| 440| 429| 491| 2,773| || Female| 885| 853| 926| 851| 797| 988| 5,300| || **Total**| 2,682| 3,534| 4,635| 8,301| 3,517| 5,526| 28,195| ## Analysis ### Determining the dimensions of variation Metrics were taken from Labov et al.’s ([2013][9]) study examining sound change over the last century in Philadelphia’s white population. * **EY-Raising**: Front Diagonal `F2-2*F1` * **Canadian Raising** (AY-Raising): `F1` * **OH-Lowering** (/ɔ/): `F1` where *F1* is the first formant of acoustic resonance of the vocal tract and *F2* is the second formant. ### Visual inspection, smoothing, and pairwise significance tests All data visualizations are conducted using the ggplot2 package in R (R Core Team, 2017; Wickham, 2009). Visual examinations of the data by year of birth make use of generalized additive smooths. A generalized additive smooth is a combination of smooth functions that provides the closest fit to the data given the parameters assigned. ### Generalized Additive Mixed Models (GAMMs) The generalized additive smooths used in data visualization are themselves a result of generalized additive modeling (cf. Fruehwald, 2017; Winter & Wieling, 2016; Wood, 2017), which was used to test for statistical significance in the apparent time data as well. Rather than assume a linear relationship between the dependent and independent variables, as is the case in linear regression (including its mixed effects analogues), generalized additive models fit a series of smooth functions to the dependent variable to account for a variety of patterns that may exist between the two. ### Measuring vowel merger or maintenance The /ɔ/ vowel is reported to be undergoing reversal in Philadelphia, lowering in younger speakers to avoid stigmatization (Labov et al., 2013; Prichard & Tamminga, 2012). However, it has been claimed that this process does not introduce a merger of /ɔ/ with /ɑ/, which is a rapidly advancing change in-progress across the United States (the CAUGHT-COT merger; cf. Labov et al., 2005). #### Metric: Pillai score * Derived from multiple analysis of variance (MANOVA) on ordered pairs of F1 (vowel height) and F2 (vowel frontness; cf. Babel, McAuliffe, & Haber, 2013; Berry & Ernestus, 2018; Hay, Warren, & Drager, 2006; Labov et al., 2016; Nycz & Hall-Lew, 2013). * These ordered pairs are predicted by the vowel label, allowing the model to indicate the degree to which the vowel label accounts for the vowel’s position in F1xF2 space. $$ Pillai\ Score: \sum_{i=1}^{p} \frac{λ_i}{1+λ_i} $$ In this case, Vowel Category is the only independent variable, p = 1, and the Pillai score ranges from zero to one . * Scores close to zero indicate that the vowel category label accounts for very little variance in F1 and F2 in the data, and thus imply that the categories are strongly merged. * Scores close to one indicate that the vowel category label strongly explains the observed variance in F1 and F2, and hence provide evidence that the two categories do not overlap. In other words, the Pillai score can be interpreted as an approximation to the degree of separation between two vowels in `F1 x F2` space. * Any speaker with fewer than 20 tokens of /ɔ/ and /ɑ/ in either speech style was excluded from the Pillai analysis for that style. [1]: https://osf.io/2qzgk [2]: https://audiobrains.com/download/beyerdynamic/stagemicrophone/Touringgear/headset/TGH54c/DAT_TGH54c_EN_A3.pdf [3]: https://files.eric.ed.gov/fulltext/ED250938.pdf [4]: https://osdoc.cogsci.nl/ [5]: http://nmcode-switching.la.psu.edu [6]: https://github.com/JoFrhwld/FAVE [7]: https://en.wikipedia.org/wiki/ARPABET [8]: https://osf.io/3ab9u [9]: http://muse.jhu.edu/article/503024