Outline

1. Corpus sociophonetics, speaking and articulation rate, previous work, research questions

2. Data collection from YouTube, transcript files, corpus creation

3. Calculation of articulation rate

4. Spatial autocorrelation: Getis-Ord G^*_i for regional analysis, urban-rural diffences

5. Caveats, summary, future outlook

• New method for the calculation of articulation rate from automatic speech-to-text transcripts
• Investigation of articulation rate vs. location, articulation rate vs. locality population
• Mapping with local autocorrelation statistics

Slides for the presentation are on my homepage at https://cc.oulu.fi/~scoats

Corpus sociophonetics

Prosodic features increasingly being considered as bearing indexicality in the same manner as (e.g.) lexical or grammatical variables (Ray & Zahn, 1999; Kendall, 2013)

Attitudes about regional or urban-rural differences in speech temporality are common in the U.S. (Preston, 1989, 1999; Roach, 1998)

Faster speech can be associated with

• Competence, intelligence, and expertise (Smith et al., 1975; Street & Brady, 1982; Thakerar & Giles, 1981)

• Persuasiveness (Apple et al., 1979; Giles & Powesland, 1975, Miller et al., 1976)

• Attractiveness (Street et al., 1983)
  

compared to slower speech

Example video (slow talker)

Example video (fast talker)

Speaking rate and articulation rate

Speaking rate: Sum of units of speech (e.g. phones, syllables, or words) divided by total utterance time

Articulation rate: Sum of units of speech divided by total utterance time, omitting pauses between segments of unbroken speech

• Pause duration has been shown to vary (Goldman-Eisler, 1961), also according to demographic and regional parameters (Clopper & Smiljanic, 2011, 2015)

• In this study articulation rate, measured in σ/sec., is compared

Factors that can affect articulation rate

• Type of speech: Reading, monologue, conversation

• Conversation: Interlocutor familiarity, topic under discussion (Yuan, Liberman & Cieri, 2006)

• Utterance-internal considerations (Byrd & Saltzman, 1998; Yuan, Liberman & Cieri, 2006; Oller, 1973)

• Anatomical, physiological, or neurological parameters (Tsao & Weismer, 1997; Tsao, Weismer & Iqbal, 2006)

• Demographic, social, or regional identity (Byrd, 1992, 1994; Jacewicz et al., 2009, 2010; Kendall, 2014)

Previous work on regional variation in speech rate in the US

Trend evident, but

• Low granularity
• Non-naturalistic data
• Small samples

Research questions

• Can we confirm the previous inferences on the basis of much larger data sets?

• Are there differences in the temporality of urban/rural speech in the United States?

Data source

• Naturalistic language data on YouTube

• Beginning in 2009, English-language videos accompanied by automatically generated speech-to-text captions (Google, 2009) with individual word timestamps

• Recent advances in neural-network-based speech-to-text transcription increase transcript accuracy (Chiu et al., 2018)

• High audio fidelity, standard language = accurate transcript = accurate word timings

• Word timings can be leveraged to calculate articulation rate

Example video

.vtt file

Focus on US local government/civic organization meetings

Public meetings of elected representatives at town/city/county/state level: advantages in terms of representativeness and comparability

• Speaker place of residence (cf. videos collected based on place-name search alone)

• Topical contents comparable

• Communicative contexts comparable

• Audio quality often high

Data collection and corpus

Script to search YouTube API for channels:

• Substrings county of, city of, municipal, town meeting, city council, county supervisors, board of supervisors, government, and official government + names/abbreviations 50 U.S. states or names of the 312 municipalities and 100 counties by population in the United States + corresponding state names/abbreviation
• E.g. county of Alabama, city council CA, official government Chicago, Illinois, official government Los Angeles County, California
• 1,680 channel matches, remove false positives and duplicates → 579 channels
• Download all captions files (53,743) in .vtt format
• Script to extract text and timings and assign exact latitude-longitude coordinates
• Filtering: Remove channels with aggregate transcripts < 1,000 words, geographically delimit to 48 contiguous states
• 49,345 videos; 28,166.77 hours of video; 235,824,795 words (Coats, 2019)

Calculation of articulation rate

• Words and their timing metadata are arranged sequentially in a captions block

• Timings within a block do not overlap

• If speech is continuous, approximate word durations (and articulation rate) can be calculated by subtracting a word's start time from that of the following word.

BUT: the annotation does not explicitly indicate word ending, so silences within utterances or between utterances by different speakers = loooooong words

AND: Utterance-initial words and those after long pauses are assigned a length of one second

Calculation of articulation rate 2

Calculation of articulation rate 3

Articulation rate: .610 σ/sec. = Not accurate

Calculation of articulation rate 4

Articulation rate: 5.26 σ/sec. = Reasonable

Calculation of articulation rate 5

• Filter out word tokens with long durations (utterance-initial words and words spoken immediately before or following longer pauses)

Intra-utterance Continuous Articulation Rate

• Articulation rate, in syllables per second, for all word tokens in a captions file whose sequential duration is less than 1 second

• Validation of method: Use Praat script speechrate (De Jong & Wempe, 2009) to calculate articulation rate directly from audio files for random sample of 20 videos from corpus, Pearson's $r = 0.83$

Regional analysis

• For each channel location: Calculate the mean intra-utterance continuous articulation rate, based on all videos from that channel

• Use the local spatial autocorrelation statistic Getis-Ord G^*_i to infer large-scale patterns of difference or similarity

• Map the variate into a Voronoi tesselation

Getis-Ord G^*_i  (Ord & Getis, 1992; Getis & Ord, 1995)

Local spatial autocorrelation statistic used in geography and recently in dialectology (e.g. Grieve, 2016)

For each point in spatially distributed data: Positive value ➜ in a cluster of high values, negative value ➜ in a cluster of low values, zero ➜ not in a cluster

$$G_i^* = \frac{\sum_{j=1}^{n} w_{ij}x_j- \bar{x}\sum_{j=1}^{n} w_{ij}}{{\sqrt{\frac{\sum_{j=1}^{n} x_j^2} {n}-\bar{x}^2}}\sqrt{\frac{n\sum_{j=1}^{n} w_{ij}^2 -{(\sum_{j=1}^{n} w_{ij})^2}}{n-1}}}$$

Result is a standard deviate (significant at $p = 0.05$ for $G_i^*\geq\pm1.645$)

Regional analysis https://stcoats.github.io/artic_rate_new.html

Regional findings

• Spatial autocorrelation analysis suggests lower articulation rates in the South (Mississippi, Alabama, Tennessee) and higher rates in the Upper Midwest (Wisconsin, Minnesota, Dakotas), the Mountain West, and parts of Florida

Urban-rural difference

Caveats

• Many different video genres represented in channels (not only meetings)
• Meetings of local government not representative of speech in general?
• Measure dependent on quality of transcript ( = quality of audio)
• Large degree of variation within small regions (patterns only emerge using spatial autocorrelation statistic)

Summary and outlook

• Large corpus of automatic speech-to-text transcripts from YouTube channels of local governments
• New method to calculate articulation rate from word timing information
• Spatial autocorrelation/visualization shows that
  • Southerners speak slightly more slowly
  • People in cities speak slightly faster

• Indexicality of articulation rate
  • Look in the transcripts for lexical items known to index regional identity (dialect words), regress frequencies with articulation rate
• Pause frequency and duration analysis
• Automatic identification of higher-quality (more accurate) transcripts
• Automatic annotation of speaker variables?

Thank you!

References

References II