Articulation Rate in American English in a Corpus of YouTube Videos

Abstract

Previous studies of the temporal organization of speech in American English have found differences in speaking or articulation rate according to speaker dialect or location, but small sample sizes and incomplete geographic coverage have limited the generalizability of the findings. In this study, articulation rates in American English are calculated from the automatic speech-to-text transcripts of more than 29,000 hours of video from local government and civic organization channels on YouTube from the 48 contiguous US states, containing more than 230 million individual word timings. Two questions are considered: are there regional differences in articulation rate? And do urban speakers articulate faster than rural speakers? The study presents several methodological innovations: first, it identifies a genre of regional speech suitable for interregional comparisons (meetings of local governments or civic organizations). Second, it introduces a new method for the calculation of articulation rate using cue and word timestamps from caption files. Third, it leverages US Census data to correlate the articulation rate with population for a large number of localities. The study shows that, in line with previous studies, Southerners articulate slower, and Americans from the Upper Midwest more quickly. In addition, there is a small but positive correlation between population size and articulation rate. Articulation rates are mapped using a measure of local autocorrelation.

A database of North American double modals and self-repairs from YouTube

“…YouTube videos can be rich sources of data for linguistic analyses, whether qualitative (e.g., Androutsopoulos, 2013) or quantitative (e.g., Coats, 2020), but communication on the platform is not limited to the utterances recorded in the site's video content. CoNASE and the DMSR provide direct access to video and audio content via metadata associated with each transcript file.…”

Computational sociophonetics using automatic speech recognition

“…Data can also be gathered from publicly available Internet resources like social media sites. For example, Coats (2020) used subtitles from 29,000 h of YouTube videos and 200 min of video downloaded using the youtube‐dl programme to study geographical variation in speech rate between Southern English and Standard American English (Amine et al., 2011). This strategy might eliminate the need for ASR because these videos are already transcribed.…”

Articulation rates’ inter-correlations and discriminating powers in an English speech corpus

“…Jessen, 2007): for example, Tauroza and Allison (1990) report a mean canonical syllable rate of 4.3 sylls/sec (to be precise, 260 sylls/minute) for their sample of conversational speech, but appear to have quantified speaking rate (including pauses) as opposed to articulation rate. Moreover, it is clear from variationist research that there is considerable variation in speech tempo within language varieties delimited as broadly as 'British English' (see Clopper & Smiljanic, 2015;Coats, 2019;Jacewicz et al, 2010;Kendall, 2013;Kowal, Wiese, & O'Connell, 1983;Quené, 2008). The corpus of Lee and Doherty (2017) seems comparable in design to the DyViS database, although its speakers are simply described as 'Irish English'.…”