Using social media as a source of analysable material in phonetics and phonology – lenition in Spanish

1.1 Social media as a source of data for phonetic analysis

Social media provides an enormous source of language data produced in real-life situations and without the oversight of a researcher. On the one hand, this leads to uncontrolled samples where the produced sentences are not designed to contain the objects of interest, which makes it more difficult to get enough data to investigate the desired phenomena; larger corpora may be needed to elicit what the researcher is looking for. On the other hand, we get the language as it is used on an everyday basis, in real-life situations, and with real interlocutors, without hypercorrection or inhibition produced by a laboratory setting or the presence of a recording device.

Unconstrained language productions have to date been the object of interest mostly of text linguists (see e.g. Sun et al. 2021 for a review; also Baldwin et al. 2013; Rüdiger and Dayter 2020). However, it is also possible to use speech recordings produced via social media either for sociolinguistic or, more specifically, for (socio)phonetic analysis. A major concern in the latter case is recording quality, as only high quality recordings are suitable for a more in-depth acoustic analysis. This topic has been examined fairly recently in the context of remote data collection (see especially Freeman and De Decker 2021; Zhang et al. 2020, 2021). Zhang et al. (2021), for instance, compare different devices and applications used either on a computer or on a smartphone, including video conferencing, for remote data collection, pointing to the advantages and disadvantages of this method. While such data collection proved useful, several data quality issues, such as formant and F0 distortions, need to be taken into account. However, not all types of phonetic analysis require the absolute reliability of these metrics, which is of consequence for the present study. Analysing relative duration and intensity is (one hopes) a lot simpler and less sensitive to data quality issues. Of relevance here is that these are the two most important variables in studying lenition. At the same time, there is an important distinction between remote but controlled data collection and naturalistic recordings made over social media. While remote data collection is a conscious process which involves a task (usually reading or repeating sounds or sentences), social media recordings give access to more authentic productions because they are not made for the purpose of phonetic analysis. Rather, the data are later repurposed for speech analysis (with the speaker’s consent). This is a major advantage in studying lenition, where we are dealing with optionality and variation, and we often want to get insight into the rates of sound changes not only inside, but also across, words.

In this context, I would like to show that sufficient quality can be obtained from social media recordings at least for some types of phonetic/phonological analysis. Thus, in the subsequent sections of the paper, I present the results of a comparative analysis of recordings obtained via a popular social media application, WhatsApp (https://www.whatsapp.com), and recordings obtained in an experimental setting. As will be argued in Section 5, naturalistic recordings give us insight into the actual productions of speakers in the course of peer-to-peer communication and into the extent of lenition in this context, which is much greater than when measured in a controlled study.

3.1 Data under analysis

The first corpus of data consists of 39 sentences each containing one instance of postvocalic /p t k/ and produced twice by five native speakers of Gran Canarian Spanish (one female, four males) aged 23–24. Each sentence consisted of the phrase He comprado cinco ‘I have bought five’ followed by a noun phrase consisting of a noun denoting a container or an object (the target word) and a prepositional phrase, such as cubos de basura ‘trash bins’. The preceding sound was [o], whereas the target segment was [p], [t], or [k], in which postvocalic lenition was expected. There were 13 p-initial, 13 t-initial, and 13 k-initial words (see Broś and Lipowska 2019 for details).^[3] The recordings were made using a Zoom H4N digital recorder and a Shure SM10a headworn microphone.

The second corpus consists of spontaneous recordings made by the speakers themselves in the course of their interactions with each other via WhatsApp. The preliminary corpus consisted of 25 short voice messages. While some of these recordings can be useful for sociolinguistic or impressionistic phonological analysis, they were deemed unsuitable for acoustic analysis due to insufficient sound quality. Only a subset of messages were recorded in a silent setting. Those recorded outdoors are too noisy to allow for a reliable analysis of voice distinctions or the intensity profile of the sounds of interest. Nevertheless, the remaining 16 recordings were of sufficient quality to allow for acoustic measurements (see Table 1).

3.2 Acoustic measurements

To determine the amount of voicing, I followed the procedure used by Broś and Lipowska (2019), that is, I deemed a given sound partially voiced if more than 50% of the sound duration showed voicing on the spectrogram and fully voiced if pulses were present throughout the sound following a visual inspection taking into consideration the voicing bar and the voice report from Praat (Boersma and Weenink 2021).

To assess the degree of aperture of the sound in question, that is, relative degree of lenition, I measured the minimum intensity of the consonant and the maximum intensity of the following vowel to calculate the intensity difference in decibels. It is a general rule in lenition literature that the more lenited the consonant, the higher its minimum intensity and hence the smaller the intensity difference between the consonant and the following vowel (Hualde and Nadeu 2011; Hualde et al. 2011; Parrell 2010). This principle has been taken further to assume that lenition in intervocalic position leads to the greater continuity of the speech signal by making the consonant more similar to the flanking vowels (Katz 2016; Kingston 2008).

According to the literature, the duration of the sound is correlated with the degree of weakening (see e.g. Cohen Priva and Gleason 2020), thus I also measured relative sound duration, following Dalcher (2008) and Broś and Lipowska (2019), as a ratio of total consonant duration to total VCV sequence duration.

Other variables taken into account in the analysis were: the presence of burst and the presence of formants. Given that bursts are produced only in maximally constricted consonants in which complete closure must be maintained for at least 20–30 ms (Shadle 1997), lenition can lead to burst-less stops, either voiced or voiceless (e.g. Dalcher 2008). Also, weakened consonants are less constricted in terms of muscle tenseness and proximity of the articulators (Vennemann 1988), which results in changes in manner (i.e. approximantization). As mentioned above, voiceless stops in the dialect have a tendency to be weakened in precisely this way, which results in the presence of formants on the spectrogram.

3.3 Statistical methods

The analyses presented below were all conducted in R 4.1.2 (R Core Team 2020) with the package lmer4 (Bates et al. 2018). Two linear mixed-effects regression models were estimated using the function lmer(), each run for one of the two numerical variables – intensity difference and relative consonant duration – as the dependent variable. Additionally, mixed-effects binary logistic regression models were built for the following binary variables: burst, voicing, and formants. This was done using the glmer() function with binomial family and the logit link function. In each model, setting (lab or social media) was used as a fixed factor and there were random intercepts for speaker, word used, and, additionally, place of articulation (labial, coronal, dorsal). A final set of models was then built with intensity difference as a dependent variable and all the other variables as fixed factors (see Section 4.2), from which an optimal fit was chosen. Full random structure was not employed due to problems with model convergence, thus no random slopes were included.

4.1 Descriptive statistics

A total of 670 observations from five speakers were analysed. The mean number of observations per speaker was 134 (SD = 27.62). At the same time, the number of target consonants produced in each setting was similar (377 in the lab setting vs. 293 in the social media app), which makes the samples comparable, with all speakers producing 68–78 target sounds in the lab setting and slightly fewer sounds, 59 on average, in the social media files, although it must be noted that the latter mean is as high as it is because of one speaker (see Table 1).

Turning to the sounds that were classified as voiced in the data, they were a minority of productions in the lab setting (43.8%) but a majority in the social media data (76%). This change is dramatic and the general reversal of voicing versus no voicing holds for each speaker (see Figure 2).

Figure 2:

Proportion of voiced versus unvoiced productions by speaker and setting.

Quite interestingly, we can see in Figure 2 that while there are substantial inter-individual differences between speakers in the lab setting (for instance, Speaker 3 has very little voicing in general), all speakers seem to be quite uniform in the percentage of voicing in a naturalistic setting with only Speaker 1, a heavy voicer in the lab setting as well, seeming to differ somewhat. This suggests that speakers in the same age range speak in a similar fashion, with similar rates of lenition, but their strategies pertaining to supervised speech differ. Some of them perhaps became more nervous or have a greater tendency for hypercorrection, while others may not have been as intimidated by the way in which they were being recorded. This issue should be pursued further in a separate study, however, due to the small number of speakers considered here.

As for the intensity difference of the produced sounds, this differed depending on the setting (see Figure 3). In the lab, the speakers tended to produce stops with a greater intensity difference (M = 19.8, SD = 7.99), that is, with less lenition compared to the social media setting (M = 14.9, SD = 6.83). The same tendency was observed in the relative sound duration. As shown in Figure 4, sounds were generally longer (M = 0.36, SD = 0.08) in the lab setting compared to the social media setting (M = 0.32, SD = 0.09), which also indicates less lenition.

Figure 3:

Intensity difference (in dB) by subject and setting.

Figure 4:

Relative consonant duration (ratio) by subject and setting.

As for other markers of lenition, the number of sounds produced without a burst, regardless of voicing, was greater in the social media setting (60.4% vs. 28.6% in the lab); the number of sounds produced with formants visible on the spectrogram was also greater (31.4% vs. 24% in the lab); see Figure 5.

Figure 5:

Presence of burst (left) and formants (right) by setting, in percentages.

It is worth observing that the overall intensity differences between sounds classified as approximants, voiced stops, and voiceless stops are very similar in the two settings and in line with previous literature (e.g. Broś et al. 2021), according to which underlying /p t k/ produced as approximants have the smallest intensity difference (around 5–12 dB), those produced as voiced stops are in the mid-range (around 15–20 dB), and those which surface as voiceless stops have values closer to 22–24 dB. Although the ranges of values are slightly different in the case of social media compared to the lab setting, the proportional changes between them are comparable. Since lenition and gestural overlap are greater in the social media context, this may translate into overall lower values of intensity for all surface pronunciations of underlying /p t k/ and hence lower than expected values in voiced stops, for instance (see Figure 6). Taking all of the above into consideration, however, there is reason to believe that the data coming from social media are reliable in terms of capturing the intensity of the produced sounds.

Figure 6:

Intensity difference (in dB) by sound category and setting. Black circles inside the coloured boxes show the means.

4.2 Statistical models

The effect of setting was significant in all the models. Intensity difference was much smaller in the social media recordings compared to the lab setting (t(128.05) = −8.3, p < 0.001). Sound duration was shorter in the social media context (t(158.81) = −5.55, p < 0.001). Also, there was an increased probability of voicing (z = 8.04, p < 0.001), a lower probability of producing underlying /p t k/ with a burst (z = −7.48, p < 0.001), and a higher probability of producing a sound with visible formants (z = 2.34, p = 0.019) in the social media recordings compared to the lab setting.

The additional model built for the data using intensity difference as a dependent variable (see Section 3.3) included all the remaining variables as fixed factors and first degree interactions between setting and each of burst, formants, and relative consonant duration, and between voicing and burst. The reason for running this model was that intensity changes are perhaps the most common marker of lenition according to previous studies (Cohen Priva and Gleason 2020; Katz 2016). Thus, I wanted to see how the degree of lenition is moderated by the remaining phonetic factors, and by the setting in which a given sound is produced. I used the lmer() function with a maximal model and then compared it with simpler models using anova(). As a result, a final best fit was selected with fixed factors as above and the following interactions: burst*setting, voicing*setting, relative duration*setting. Setting (F = 9.87, p < 0.01), relative duration (F = 31.7, p < 0.001), and formants (F = 39.76, p < 0.001) resulted significant, while burst was at the statistical tendency level (F = 3.69, p = 0.55). As for the interactions, all of them resulted significant (see Table 2 for the summary of the results). The effect plots (Figure 7) show that sounds produced with a burst have a larger intensity difference with respect to the following vowel in the social media setting compared to the sounds produced without a burst. As for voicing, when it was produced, it was correlated with a smaller intensity difference compared to voiceless in both settings in a quite similar way. Finally, greater relative sound duration was correlated with a smaller intensity difference in the social media setting compared to the lab and a positive correlation between the two variables only ensues in the lab setting (see Table 2 for the estimates).

Figure 7:

Effect plots for interactions in the final model. Predictor effect plots made with the effects() function in R provide a graphical summary of a fitted regression model with linear predictors and constitute a good way of summarizing the results beyond coefficient tables. Here, the error bars and the shaded areas correspond to the 95% confidence intervals for the fitted values. Note that when these overlap (as on the first plot), it signifies greater data variability and suggests that the plotted difference may not be significant.