1.1 Stop lenition in Canary Islands Spanish

Canary Islands Spanish is well-known for its extensive consonantal weakening, including voicing, approximantization, debuccalization and elision (Trujillo, 1980; Oftedal, 1985; Almeida & Díaz Alayón, 1988). In the case of voiced stops, the dialect follows the pattern typical of many other varieties of Spanish, i.e., approximantization in intervocalic position, with a substantial rate of elisions, especially within words, similar to that found in Caribbean dialects (Lipski, 1994; Hualde, 2005). Apart from that, voiceless stops undergo a parallel process of weakening in the dialect, by which /p t k/ are realized as partially or fully voiced, and sometimes even approximantized segments. The prevalence of /b d g/ weakening is much greater and amounts to over 95% in intervocalic positions, while /p t k/ tend to weaken with a probability closer to 50% (Broś et al., 2021, p. 21). The latter process has been shown to be suppressed to some extent in lab conditions. For instance, a study comparing /p t k/ weakening between lab recordings of sentences repeated after a native speaker and recordings made over an instant messaging app shows a substantial increase in the frequency of voicing in the latter case (Broś, 2023). This suggests that /p t k/ weakening can be suppressed by speakers and subject to the observer’s paradox (see Section 1.2). Some examples of stop weakening found in Canary Islands Spanish can be found below.

As for the parameters used to assess weakening degree and frequency in stop lenition in Spanish, researchers have used visual inspection of the spectrograms to annotate partial or full voicing, and presence or absence of a burst—in the case of /p t k/, to compare weakened and unweakened stops, as well as presence or absence of formants in both /p t k/ and /b d g/ weakening to mark approximantization (e.g., Dalcher, 2008). Furthermore, constriction and consonant duration have been used to determine the amount of weakening in several studies (e.g., Dalcher, 2008; Hualde et al., 2011). Finally, perhaps the most robust acoustic parameter used as a proxy for consonantal weakening in Spanish is some form of relative intensity of the target segment compared to the flanking vowels. This can take the form of an intensity ratio (Ortega-Llebaria, 2004; Colantoni & Marinescu, 2010; Carrasco et al., 2012), maximum velocity (Hualde et al., 2011) or intensity difference (Hualde et al., 2010; Parrell, 2010) and has been shown to correlate well with articulatory data measuring the degree of consonantal aperture (Parrell, 2010, 2011). Since intensity difference has been used to reliably show stop lenition in Spanish in numerous works, including previous work on the same dialect (e.g., Broś et al., 2021), I adopt this measurement in the present study.

1.2 Data collection methods and the observer’s paradox

The observer’s paradox within linguistics was first defined by William Labov (1972, p. 209) as a situation in which we want to know how speakers talk when not supervised by a researcher, yet this knowledge can only be gained by systematic observation. Thus, getting naturalistic data from speakers is a challenge for the researcher and can be exacerbated further by the need to get a controlled sample that allows a reliable (statistical) analysis of the collected data. The latter usually takes the form of a lab study or a similar research design. However, even when gathering data in the course of fieldwork, including when we speak the language or dialect in question, recording everyday speech is often difficult. This is because speakers tend to react to the presence of the researcher by suppressing certain linguistic features, using hypercorrection or a more formal register. The social setting (or recording environment) itself can also alter the behavior of the speaker (Wagner et al., 2015). The presence of a recording device, sitting in a lab or recording studio, the speaker’s awareness that the recording will be heard or analyzed by someone, and the nature of the task (e.g., reading or repeating words and phrases), can all drive (subconscious) changes in production. These issues have been addressed in the literature using such notions as (conscious) ‘attention to speech’ (which is minimal in the most informal styles, i.e., the so-called vernacular, Labov, 1978), ‘audience design’ (i.e., the presence and roles of interlocutors or overhearers of a speaking situation, Bell, 1984) and ‘researcher as audience’ (Wilson, 1987), among others. In the latter case, the researcher may play different roles, depending on the research method, and may even be a member of the audience indirectly, via the recording device, which acts as a proxy in the minds of study participants. Thus, speakers may be affected during data collection in a variety of ways and determining what is ‘natural speech’ may be extremely difficult without a proper comparative approach.

Technological innovation can help us at least partially remedy the situation. For example, we can gather speech samples remotely with the use of the speakers’ personal devices, which brings us closer to the everyday situations experienced by participants. Furthermore, by removing ourselves from the process of speech production, i.e., resigning from supervising data collection, we may even escape the observer’s paradox altogether.

Due to the COVID-19 pandemic, many researchers turned to alternative ways of collecting data, and a number of new studies on the matter were published as a result. Thanks to these initiatives within the field of phonetics and phonology, we now know that some types of remote recordings are better than others. For instance, Freeman and De Decker (2021a) show that video conferencing apps such as Zoom may be a good way of collecting data for the analysis of general vowel arrangement properties, but researchers looking at small formant differences and vowel overlap, as well as nasalization, should proceed with caution. Some distortions and excessive within-data variation can pose a problem to acoustic analyses. Zhang et al. (2020) compared recordings made over Zoom with those produced over smartphones using lossless format apps such as Recorder or Awesome Voice Recorder, showing that smartphone use is more reliable and can serve for analyzing some parts of the speech signal, e.g., prosody. They also note that irregular waveforms and filtering artefacts resulting in temporal misalignments can be produced in the course of conversion of recordings made on Zoom, and that remote recordings are, in general, more reliable when analyzing vowels than tones. In a similar study, Zhang et al. (2021) show that, unlike smartphone recordings, Zoom recordings are burdened with sudden intensity drops, and thus smartphones may be more suitable for collecting data aimed at providing some types of phonetic analysis (e.g., f₀ and formant measurements).

The potential reliability of homemade recordings for F1 and F2 analysis was also confirmed by Freeman and De Decker (2021b) in a study that simulated the use of personal devices by speakers. Importantly, they recommend that laptops be used whenever possible, as they are more reliable than iPads or smartphones. Contrary to that, however, Sanker et al. (2021) make slightly different recommendations after comparing various types of devices and several types of recording software in a simultaneous recording setup. Their analysis shows that different types of distortions are produced when recording on an iPad, smartphone or laptop without an external microphone. They also point to Facebook Messenger, Skype and Zoom as software options that affect many aspects of the speech signal.¹ In their investigation, Sanker et al. go beyond vowel productions and compare many different phonetic parameters that play a role in acoustic analysis. These include center of gravity (crucial for the analysis of fricatives), harmonics-to-noise ratio (important when looking at voicing and noise parameters in the data), spectral tilt (that can involve intensity and f₀ distortions), and sound duration, among others. The authors of the study suggest that care be taken in making sure that possible distortions are taken into account, together with such issues as file format (lossless vs. lossy), sampling rate (preferably 44.1 kHz/s), filtering options and filter-related artefacts, as well as recording comparability (preferably, limiting recordings to one session and comparing the same recording types) when conducting remote fieldwork or using self-recordings.

As for the use of smartphone recordings in phonetic research, not much has been published so far beyond the controlled methodological studies mentioned above and looking at phenomena related to vowel production and nasalization. It is worth noting that the general suitability of smartphones as recording devices comparable to traditional means was demonstrated earlier, in 2011 (De Decker & Nycz, 2011). However, researchers should proceed with caution given the findings mentioned above. Moreover, using alternative devices instead of traditional fieldwork recorders is not the same as collecting data remotely. The building block added by the pandemic concerns the latter issue.

Yet another subject for linguists to take up is unsupervised recording, on which even less research has been done to date.² Gittelson et al. (2021) report one such study that looked at crowd-sourced data gathered via a dedicated phone application. The participants had to engage with the app and predetermined elicitation protocols, but they used their phones independently of third persons in order to take part in the study. The project involved an investigation of the sociolinguistic aspects of nonmodal phonation in English and, apart from helping researchers reach more participants than via traditional means, it showed promising results related to inter-speaker variation. According to the authors, the ability to collect big data from a more diverse sample helped discover new factors influencing the use of non-modal phonation. An earlier study by Hall-Lew and Boyd (2017) studied sociophonetic variation in sibilant production based on a small sample of self-recordings made by four speakers of American English. By comparing interviews and controlled speech samples produced under supervision with the self-recordings, the researchers concluded that self-recordings are suitable for intra-speaker comparisons and can give us insight on the range of productions that may not have been captured otherwise. They also note, however, that such recordings may be produced with a range of styles across speakers, which should be taken into account in inter-speaker comparisons.

Finally, a recent study by Broś (2023) taps into intra-speaker differences in the production of voiceless stops in Canary Islands Spanish depending on whether the speech was supervised and recorded in a lab setting or came from unsupervised recordings taken from an IM (instant messaging) application (WhatsApp). Apart from looking at the suitability of such recordings for phonetic analysis, the most important question was whether and to what degree native speakers suppress /p t k/ lenition in lab speech. Since WhatsApp samples were examples of natural speech and participants had no expectation that their recordings would undergo third-party analysis, Broś assumed that these samples represented everyday speech spoken by representatives of the dialect. The study showed that the two tested recording types are suitable for intensity-based comparisons and phonetic annotation. Based on the voicing and other parameters of the target segments, a discrepancy between the recording types was discovered, with significantly more lenition taking place when speaking via IM and, importantly, a levelling effect between speakers. While the number of weakened sounds and the degree of weakening differed substantially from speaker to speaker in lab speech, all speakers produced the same amount and type of lenitions when talking to their friends over the phone application.³ Overall, the results of the study show what we potentially lose when collecting lab speech and how this affects our generalizations concerning the way in which a given speech community talks. Potentially, we may significantly underestimate the frequency of a given process or overlook some contexts of application of a phonetic or phonological rule. This is especially important when studying language variation and change.

In the present study, I use recorded speech produced over WhatsApp as a baseline condition that, presumably, best corresponds to the actual speech of the inhabitants of the Canary Islands. Apart from that, I test a guided self-recording method as a possible means towards reaching this baseline condition and getting more natural speech compared to data produced during a field experiment consisting in reading sentences. Finally, I use fieldwork recordings of spontaneous speech to disentangle recording type from task type. With such a multivariate comparison, we can potentially work out an optimal method of data collection to discover the speaking habits closer to a community’s everyday speech behavior.

1.3 The current study: Goals and hypotheses

In this study, I use bilabial stops for two reasons. First, only /p b/ were used in the experimental condition that I take to compare recording types (supervised vs. unsupervised) and tasks (scripted sentences vs. spontaneous productions). The second reason is phonetic. In general, labials are more suitable for across-task comparisons, given that they are less susceptible to the effects of vowel coarticulation. Moreover, it has been suggested that labial stops are the most compatible with voicing compared to other stop series (Ohala 1983, p. 195) and while their constriction is the longest, their voicing durations can be easily prolonged by passive or active vocal tract expansion. In the same vein, Maddieson (1984, p. 36) states that “voicing is more readily sustained in a bilabial plosive than in any other,” because the air can keep flowing into the oral cavity for a longer period before air pressure equals subglottal pressure when the closure is far away, i.e., at the lips. This facility in voicing and the challenging task of maintaining voicing in other stops, especially velars, is supported by the data from the dialect. Broś and Lipowska (2019) have shown that while all stops are voiced by Gran Canarian speakers in post-vocalic position around 45% of the time, /p/ is the most likely to be voiced, while /k/ tends to be shortened more often, and/or approximantized instead.

As the main goal of this study is to compare different recordings and establish the viability of self-recordings for studying lenition, recording type becomes a crucial variable. However, research on the differences between self-recorded speech and controlled recording sessions in the context of particular phonetic and phonological phenomena is limited. There are only a few small-scale studies that might inform us on what is to be expected of self-recorded speech. Boyd et al. (2015), for instance, have shown that self-recordings present the most advanced vocalic changes corresponding to the California Vowel Shift compared to recorded interviews. Similarly, the study by Hall-Lew and Boyd (2017), described in Section 1.2, which analyzed /s/ productions by California English speakers, showed that self-recordings can help uncover a wider range of phonetic variants produced by a given community that surpasses what is revealed by interview data. Thus, it may be assumed that self-recordings should be the closest to natural productions made by speakers for purposes other than linguistic analysis upon researcher instruction. Assuming that IM recordings are instances of natural speech par excellence (as suggested in previous work by Broś, 2023), I would like to test the hypothesis that self-recordings will be closer to this ideal compared to experimental data, both in the acoustic lenition marker (i.e., intensity difference) and in the frequency of lenition.

Furthermore, the recording types used in this study also differ in what type of speech (or task) is used. Presumably, scripted speech in the form of sentences provided by the experimenter may induce less naturalistic productions whose features differ compared to monologues or interviews. While Martínez-Gil (2020) argues that /b d g/ approximantization takes place even in slow and careful speech, there is some experimental work that shows task effects with differences in the degree of constriction of /b d g/ between conversational speech and reading tasks in monolingual and heritage Spanish speakers (e.g., Carrasco et al., 2012; Lozano, 2021). As for voiceless stops, Lewis (2001) found a significant difference in voiceless stop productions between a word-list reading task and conversational speech among Colombian Spanish speakers. Similarly, Hualde et al. (2011) compared scripted and unscripted data, showing that spontaneous speech results in more /p t k/ voicing among Spanish speakers from Majorca. There is also some evidence from the dialect studied here: Herrera Santana (1997) tested Gran Canarian speakers and reported that when asked to read lists of words containing intervocalic voiceless stops, speakers voiced them only rarely. Given this evidence, I assume that scripted speech (read or repeated sentences) will induce less lenition compared to spontaneous speech (monologues and interviews).

Finally, this study also assessed how a person’s speech rate, which is tied to speaking styles and registers, can change depending on the type of recording used for gathering data. Since speakers use different modes of speech that are influenced by whether or not their produced sentences are scripted, and whether a third person (here, the experimenter) is directly monitoring their speech, I assume that differing speech rates will be found in the different recording samples analyzed in this study. Moreover, higher speech rates induce shorter sound durations and, indirectly, greater coarticulation, gestural overlap and/or undershoot (Cohen Priva, 2017). This may translate into more advanced lenition, as manifested by acoustic parameters (e.g., a smaller intensity difference, more voicing, etc.) and/or into a higher probability of applying lenition in a given context. Previous studies have shown that speech rate plays a role in lenition in general (Cohen Priva & Gleason, 2020), and in Spanish in particular (Soler & Romero, 1999; Hualde et al., 2011; Nadeu & Hualde, 2015; Melero-García, 2021; see also Broś et al., 2021, Section 5.5). Thus, speech rate should be included in the statistical analyses comparing tasks and recordings. I expect that higher speech rates will increase the probability and amount of voicing, whereas in the case of /b/, I expect that higher speech rates will translate into more deletions.

The hypotheses to be tested in this study are as follows.

Hypothesis 1

Speakers differ in their production of underlying /p/ and /b/ across all recording types: Self-recordings show a smaller intensity difference, i.e., more lenition, compared to the field experiment, with values closer to IM recordings.

Hypothesis 2

The frequency of occurrence of weakening depends on the recording: a) In the case of /p/, there is more voicing in self-recordings compared to the experiment and more voicing still in IM recordings; b) in the case of /b/, there is more deletion in self-recordings compared to the experiment, and even more in IM recordings.

Hypothesis 3

Speakers differ in their production of underlying /p/ and /b/ depending on task type: Spontaneous speech (monologues) will be correlated with more lenition compared to scripted speech (sentences), regardless of recording type. This will result in:

• -  a smaller intensity difference

• -  a greater frequency of voicing in /p/

• -  a greater frequency of deletion in /b/

In all cases, I assume a qualitative difference between underlying /p/ and /b/, i.e., a greater intensity difference in /p/ than in /b/, given that the former is typically produced as either a voiceless or a voiced stop and the latter – as an approximant (when not deleted).

Hypothesis 4

Since /p b/ weakening is a connected speech phenomenon and has been shown to be sensitive to the durational properties of language, I expect speech rate to modulate weakening in both segments. The higher the speech rate, the greater the probability of lenition (by voicing or deletion, accordingly) and the smaller the intensity difference.

2.1 Materials and equipment

The data analyzed in this paper come from eight native speakers of Canary Islands Spanish (six males, two females) who participated in a field experiment in 2021 (Broś & Krause, 2024). Twelve participants from that study were asked to provide self-recordings so that a comparative analysis could be performed. Only eight of them completed the task. The remaining four either refused for lack of time or were unresponsive.

The self-recording consisted of two parts. First, each participant was asked to talk about their job, recent vacation or any other topic of choice for approximately one minute. As a second task, they were asked to listen to a recording with 40 sentences and repeat what they heard. A speaker of the dialect produced the sentences for this recording. A period of silence was inserted after each sentence to give the participant enough time to repeat it. The silence was longer than the sentence itself.

To ensure data comparability, I selected two independent productions of the same 20 sentences taken from the 2021 study, containing 17 target words with an intervocalic /p/ and 13 words with an intervocalic /b/.⁴ Each target segment appeared at the beginning of a word, flanked by two vowels /a/. Examples are provided below.

Sample sentences from the study:

• La barrera estaba mal colocada y el portero no veía.

• ‘The wall was incorrectly placed, and the goalkeeper could not see’.

• La patrulla ha encontrado el ladrón en la barca.

• ‘The patrol found the robber on the boat’.

For maximal comparability, the same sentences as those used in the self-recordings were analyzed in the experiment condition. The number of segments considered, however, is greater due to the greater number of repetitions (between five and ten) produced during the experiment.

The present study included an additional sample for six participants. Since I had access to WhatsApp recordings of these persons, and such data have been deemed good examples of naturalistic speech in the dialect (Broś, 2023), I decided to use them as a baseline in the analyses. These examples consisted of recordings made in a group conversation between the participants and, in some cases, recordings exchanged with me, via WhatsApp. The participants agreed to an analysis of the recordings for the purposes of the study. To ensure maximal comparability of the results, I only considered bilabials /p b/.

Samples 1 to 3, i.e., the field experiment and self-recordings, were recorded using Audacity⁵ or WaveEditor for Android⁶ in WAV format. Initially, I had planned for all self-recordings to be made using Audacity on a laptop computer, but this was impossible for three speakers, as they do not have laptops. Thus, I instructed them to use a smartphone app instead. After some research, I tested and then recommended WaveEditor, which allows for good quality recordings in WAV format at a high sampling rate (48 kHz/s in its default setting). Audacity was used with the default settings and a 44.1 kHz sampling frequency. The participants were asked to save the files in WAV format and send them via email or a cloud drive. In principle, I wanted to make sure samples were comparable. Since the field experiment had recordings made using Audacity on a laptop computer, the intent was to use the same tools in the present study, but in a self-recording design in which speakers independently use the software, record themselves and provide the resultant files.⁷ Sample 4 consisted of recordings in OPUS format, which is a lossless file type used by WhatsApp.⁸ The recordings were then converted to WAV for analysis in Praat (Boersma & Weenink, 2022).

Finally, given the intra-speaker differences found between sentence productions and spontaneous speech (monologues), I decided to include a fifth data sample to disentangle recording type from task type. This was possible on a subset of data, given that four of the participants had earlier participated in fieldwork interviews with me. Those were conducted in the same locality as the experiment, but with a professional recording device (a Zoom H4N digital recorder with a Shure SM10a headworn microphone with a 44.1 kHz sampling frequency). Since the samples were made in the presence of the experimenter who also controlled them, they constitute a good comparative sample for a 2 × 2 (recording x task) design, in which field sentences and field monologues may be compared to self-recorded sentences and self-recorded monologues.

Thus, all in all, I analyzed five data samples:

• a)  8 field experiment recordings of read sentences

• b)  8 self-recordings of the same sentences in a repeating condition

• c)  8 self-recordings of spontaneous speech (monologue)

• d)  6 instant messaging app recordings made via WhatsApp

• e)  4 field recordings from semi-structured interviews (monologue)

All recordings were annotated in Praat. All target segments and flanking vowels were entered in the TextGrids together with the information on voicing in the case of /p/; deletion rate in the case of /b/, stress, and word in which a given segment occurred; and the sentence it belonged to. A sample annotation is presented in the Appendix.

Following previous work on Spanish stop weakening in the same dialect (Broś et al., 2021), target segments were deemed voiced based on the visual inspection of the waveforms and spectrograms when the voicing bar and pulses were present in more than 50% of the total duration of the sound. The beginning of the stop corresponded to the end of the preceding vowel, and the end was marked at the beginning of the periodic cycle of the following vowel. Approximant pronunciations were distinguished by the lack of burst, full voicing and presence of formants on the spectrogram.⁹ To delimit these segments, I followed previous work (e.g., Eddington, 2011; Hualde et al., 2011) by looking at the dip in the intensity contour from vowel to consonant and a subsequent inversion of this trend in the transition to the following vowel. When the intensity curve seemed flat and the differences in intensity between the flanking vowels were too small to reliably decide whether an approximant was produced in between, I deemed the segment deleted.

A custom-made Praat script extracted the segmental and sentence information, together with sound durations and intensity measurements in dB extracted from Praat’s intensity object. This gave me a total of 2,377 observations from eight speakers in the main analysis database (recordings 1–4). Of those, the majority came from the experiment (n = 1467), followed by self-recorded sentences (n = 526), self-recorded monologues (n = 229) and IM recordings (n = 155). In the latter case, the observations come from a subset of six speakers, as mentioned before. The fifth recording gave me an additional pool of 434 observations. Given data subsetting depending on the model, the number of observations taken under analysis is stated separately for each case in Section 3.

Table 1 lists the number of segments per participant and recording type, deleted /b/ sounds included. Note that spontaneous recordings contain fewer segments, as speakers do not necessarily use intervocalic /p b/ in their productions.

2.2 Data analysis

The compiled database was analyzed based on the annotated parameters and intensity measurements taken from Praat. Based on the raw measurements, relative intensity difference (calculated as the minimum intensity of the target segment subtracted from the maximum intensity of the preceding vowel) was added to the database. It is worth noting that intensity difference can be measured comparing the target segment with either the preceding or the following vowel. The former method was used, e.g., by Martínez-Celdrán and Regueira (2008), Figueroa and Evans (2015) and Broś et al. (2021), while the latter was used by Hualde et al. (2011) and Carrasco et al. (2012), among others. I used the difference between the minimum intensity of the target consonant and the maximum intensity of the preceding vowel, as this measurement is less sensitive to intensity differences in the vocalic segment that may result from word stress. The choice of this measurement method is especially important in the case of sentence reading or repetition, as both stressed and unstressed syllable conditions are included.¹¹

I also calculated speech rate as a mean of ten measurements taken randomly from sentences produced in each recorded sample, excluding pauses. I made sure that the measurements included sentences from various parts of the recording and excluded short phrases ridden with hesitations or prolonged vowels that might falsify the speaking rate. The resulting average number of syllables per second was established for each recording and added in statistical model formulae to test the role of speaking speed in the application of lenition in the dialect.

The database was then subjected to a statistical analysis in R (R Core Team, 2020) using the packages lme4 (Bates et al., 2018) for building models and emmeans (Lenth, 2019), for the calculation of simple effects. Descriptive plots and effects plots were generated using the ggplot2 (Wickham, 2016) and ggeffects (Lüdecke, 2018) packages, respectively.

I ran three initial models on the main database: 1) a linear mixed-effects model, with intensity difference as a dependent variable; 2) a binomial logistic mixed-effects regression model, with voicing (binary) as a dependent variable to test the probability of voicing in /p/ depending on the speech sample; and 3) a binomial logistic mixed-effects regression model, with deletion (binary) as a dependent variable to explore the probability of deletion of /b/ depending on the speech sample. After that, I built three follow-up models on a subset of data containing four speakers, with fieldwork interview data included and instant messaging excluded to disentangle recording type from task type.

Optimal models were achieved by using maximal model structures and the backward direction of the step() function (linear mixed models) or by model comparison using anova() (logistic regressions). The exact formulae of all the models are provided in Section 3. All model assumptions were checked, including normal distribution of residuals in linear mixed models. I also used pairwise comparisons available in the emmeans package to correct for multiple tests.

3.1 Intensity difference

The comparison of the main dataset (four samples) in terms of the key lenition parameter, i.e., intensity difference, shows that all recording types are comparable (see Figure 1). Crucially, there is a clear difference between underlying /p/ and /b/, as expected. Underlying voiced stops are quite similar, with values corresponding to weak approximants (~5 dB), while underlying voiceless stops are realized in the range of stops (15–25 dB). It must be remembered that these ranges vary depending on whether the segments are voiced or not. The former would get intensity difference closer to 15 dB, while tense voiceless pronunciations can get as high as 30 dB, or more, on the scale. These differences will be explored below.

Figure 1

Comparison of intensity difference values for /p/ and /b/ across the four types of recordings.

In Figure 1, we can see that IM recordings have the lowest values for /p/ compared to other recordings, which may be due to the different distributions of voiced vs. voiceless /p/ realizations, or to other factors, e.g., the distance between the mouth and the microphone. Additionally, IM recordings seem to have slightly smaller between-category differences, i.e., the distance between the realizations of /p/ and /b/. Nonetheless, this is due to the smaller intensity difference in /p/. The values for /b/ are virtually the same as those of other recordings.

To explore the in-between categories, I looked at density plots. In Figure 2, we can see three distinct sound categories in the data: voiceless stops (with the greatest intensity difference), voiced stops (in the mid-range) and voiced approximants (the lowest values). Approximants come from underlying /b/, while voiced and voiceless stops are realizations of underlying /p/.¹² The density plots also make it immediately clear that the type of task and recording do play a role in how /p b/ are produced: While there is a visible trimodal distribution of the three types of surface segments in spontaneous speech, an overlap between voiced and voiceless stop realizations is more pronounced when sentences are read aloud or repeated after a recording, which suggests that both [b] and [p] coming from /p/ are realized with similar intensity scores. Also, note that in IM recordings, there are three visible categories that are nevertheless much closer together on the intensity difference scale. Importantly, voiced realizations of /p/ have much lower intensity scores than their counterparts in other recordings (~10dB vs ~15dB). I will return to these particular results later in this section.

Figure 2

Density plots for intensity difference across recording types by voicing and underlying segment.

As mentioned in the Introduction, speech rate is one factor that may affect the amount of voicing and the acoustic parameters. The general trend is that experiment data are the fastest while, quite surprisingly, self-recordings are the slowest (see Figure 3).

Figure 3

Speech rate depending on recording type.

3.1.1 Model 1: Predicting lenition degree, as marked by intensity difference

The abovementioned observations were tested using the following linear mixed-effects model:

intensity difference ~ recording type + underlying segment + speech rate + recording type : underlying segment + recording type : speech rate + (1 + underlying segment | participant) + (1 | word)

Model 1 was run on 2,207 observations and fitted by REML, with t-tests using Satterthwaite’s method. The reference value of recording type was experiment, and the reference value of underlying segment was /p/. The results presented in Table 2 show significant effects of underlying segment and recording (where self-recorded sentences and self-recorded monologues, but not IM recordings, differed significantly from the experiment). Speech rate was not significant. There were also significant interactions between all levels of recording type in underlying /b/ compared to /p/ in experiment recordings, and a significant interaction between self-recordings and speech rate compared to the experiment.

Table 2

Summary of results of the first model.

	estimate	std. error	df	t value	p value
(Intercept)	18.962	2.419	44.764	7.838	<.001
self-recorded sentences	6.45	2.116	858.071	3.048	<.01
self-recorded monologue	–7.899	3.673	325.889	–2.15	<.05
instant messaging	3.022	4.729	564.151	0.639	.523
/b/	–11.045	1.313	11.143	–8.413	<.001
speech rate	–0.315	0.286	60.961	–1.102	.275
self-recorded sentences : /b/	–2.887	0.474	2100.838	–6.092	<.001
self-recorded monologue : /b/	–2.125	0.915	79.585	–2.323	<.05
instant messaging : /b/	2.841	1.118	116.087	2.541	<.05
self-recorded sentences : speech rate	–0.693	0.314	1035.418	–2.204	<.05
self-recorded monologue : speech rate	1.528	0.589	364.241	2.593	<.05
instant messaging : speech rate	–1.019	0.719	593.439	–1.416	.157

In the case of recording, pairwise comparisons based on the marginal means from the model show that there was a significant difference in the values of intensity difference between the experiment and IM recordings (β = 2.356, df = 232, t = 4.093, p < .001), between self-recorded sentences and IM (β = 2.741, df = 274, t = 4.525, p < .001), and between self-recorded monologues and IM (β = 3.590, df = 1005, t = 5.152, p < .001), but not between self-recordings and the experiment, or self-recorded sentences and self-recorded monologues. As for the overall variance explained by the model, R² = 0.543, most of it can be attributed to the underlying segment.

To explore the interactions, I looked at the contrasts generated using marginal means from the model. They show that all differences between recording types for /p/ productions are significant, except for the difference between self-recorded sentences and self-recorded monologues (β = –.469, df = 382.09, t = –0.611, p = .999). All differences between recording types and underlying segments were significant. As for the differences in /b/ productions across recording types, they were not significant. In other words, /b/ was produced the same way across recording types, while /p/ differed across conditions. Contrasts based on the marginal means calculated for the interaction of the two variables are presented in Table 8, in the Appendix. A graphical illustration of this interaction can be appreciated in Figure 4, while Figure 5 shows the interaction between speech rate and recording. As expected, there is a slight tendency in the direction of a smaller intensity difference with increasing the speech rate. However, self-recorded monologues defy this, showing an opposite trend.

Figure 4

Effects plot for the interaction between recording type and underlying segment (/p/ or /b/).

Figure 5

Effects plot for the interaction between recording type and speech rate.

3.2 Frequency of voicing

Descriptive statistics show that the weakening of /p/ is variable and depends on the speaking situation, as expected from the literature (see Figure 6). Self-recorded monologues had the greatest number of lenited /p/ segments, and IM recordings had the second highest number Sentences induced more unweakened pronunciations.

Figure 6

Percentage of voicing produced by the speakers depending on recording type. The data are taken from /p/ production only.

It is worth noting that speakers differ in the number of voiced sounds they produce. Figure 7 shows that the general tendency is to increase the amount of voicing in spontaneous speech compared to producing scripted sentences, regardless of how much voicing is produced in the ‘unnatural’ experiment condition, with the exception of two participants (P14 has virtually no change, and P7 shows a reverse trend). Moreover, half of the speakers go from nearly 100% to fully 100% of voicing in spontaneous speech.

Figure 7

Comparison of voicing frequencies presented by each speaker in each recording type. The figure shows data for /p/ only. Recording types: exp = experiment, self-s = self-recorded sentences, self-m = self-recorded monologue, i-m = instant messaging. Note that only six speakers provided IM recordings, hence not all speakers have a bar corresponding to this recording type.

3.2.1 Model 2: Predicting the probability of voicing

A subset of data (N = 1225) was used to predict the probability of voicing in /p/ depending on speech rate and recording type. Model 2 was fit using glmer() with the binomial function and the ‘bobyqa’ optimizer, nAGQ set to 10:

voicing ~ recording + speech rate + (1 | participant)¹³

The results of the model fit show a significant effect of recording for the monologue condition compared to the experiment but no other significant effects (see Table 3). Pairwise comparisons based on the marginal means from the model confirm the significant difference in the probability of voicing between the experiment and the self-recorded monologues (β = –1.536, z = –4.501, p < .001). They also show a significant difference between self-recorded sentences and self-recorded monologues (β = –1.894, z = –5.564, p < .001), and between self-recorded monologues and IM recordings (β = 1.158, z = 2.821, p < .05; see Table 10 in the Appendix). Figure 8 shows the effect plot from the model.

Table 3

Estimates from Model 2. Experiment was the reference level of the recording type variable.

	estimate	std. error	z value	p value
(Intercept)	–0.739	1.217	–0.607	.544
self-recorded sentences	–0.358	0.189	–1.896	.058
self-recorded monologue	1.537	0.341	4.502	<.001
IM	0.378	0.284	1.329	.184
speech rate	0.141	0.166	0.851	.395

Figure 8

Predicted probabilities of voicing depending on recording type.

3.3 Frequency of /b/ deletion

While the weakening of this segment happens nearly 100% of the time, most of it takes the form of approximantization. In some cases, however, /b/ is elided and the frequency of this process tends to vary depending on the speaker and the recording situation. Figure 9 shows that /b/ deletion is not that frequent and only goes above 25% in IM. There seems to be no effect of task type, as self-recorded monologues do not have higher rates than sentences.

Figure 9

Rate of /b/ deletion depending on recording type. The data are taken from /b/ productions only. Deletions are marked as ‘yes’, while /b/ retention by approximantization is marked as ‘no’.

3.3.1 Model 3: Predicting the probability of deletion

The above was tested via a binomial logistic regression analogous to Model 2, run on the /b/ database (N = 1152).

deletion ~ recording + speech rate + (1 | participant)

The results show a main effect of recording, with the probability of deletion being greater in all types of recordings compared to the experiment (see Table 4). The effect of speech rate was not significant. Pairwise comparisons show significant differences between the experiment and self-recorded monologues (β = –1.657, z = –5.511, p < .001), self-recorded sentences (β = –1.032, z = –4.544, p < .001) and IM recordings (β = –2.611, z = –8.701, p < .001), as well as a significant difference between self-recorded monologues and IM (β = –0.955, z = –2.820, p < .05) and between self-recorded sentences and IM (β = –1.580, z = –5.214, p < .001), but not between the self-recorded monologues and the self-recorded sentences (β = 0.625, z = 2.176, p = .130, see Table 12 in the Appendix). The effect plot showing the probability of deletion depending on the recording is presented in Figure 10.

Table 4

Estimates from Model 3. Experiment was the baseline level of the recording type variable.

	estimate	std. error	z value	p value
(Intercept)	–4.098	1.016	–4.033	<.001
self-recorded sentences	1.657	0.301	5.511	<.001
self-recorded monologue	1.032	0.227	4.544	<.001
IM	2.611	0.3	8.701	<.001
speech rate	0.208	0.142	1.465	.143

Figure 10

Predicted probabilities of /b/ deletion depending on recording type.

3.4 Disentangling recording type from task type

Given the possibility that task type may have a different effect on stop lenition than recording type, the relationship between the two was explored more closely. An additional sample of recordings from four of the participants made it possible to build a 2 × 2 × 2 model on a subset of data.¹⁴ I compared two types of tasks (spontaneous speech and sentence reading/repetition) in a more controlled setting (with the experimenter present) to the same two tasks in an uncontrolled setting (no experimenter, self-recordings).

3.4.1 Model 4: Predicting lenition degree by segment and recording type

For Model 4, I first built a linear mixed effects model with intensity difference as a dependent variable:

intensity difference ~ recording type + task type + segment + recording type : task type + task type : segment + recording type : segment + (1 + task type + recording type | participant) + (1 | word)

The model was run on 1448 observations. The results show a significant main result of segment, with no significant results of recording type or task type (see Table 5). There were, however, significant interactions between recording type and segment, task type and segment, and recording type and task type. The interaction plot in Figure 11 shows a discrepancy in the intensity difference of /p/ between the production of controlled sentences vs. monologues in recordings produced in the presence of the experimenter. However, pairwise comparisons based on emmeans only show significant differences between /p/ and /b/. No within-segment differences in either recording type or task type were reliable (see Table 13 in the Appendix). The overall variance explained by the model was R² = .384, most of which can be attributed to the underlying segment.

Table 5

Summary of results for Model 4.

	Estimate	std. error	df	t value	p value
(Intercept)	6.001	1.668	5.677	3.598	<.05
self-recording	–1.168	1.729	3.222	–0.676	.545
uncontrolled task	–0.452	2.837	3.74	–0.159	.882
/p/	7.146	1.111	139.464	6.429	<.001
self-recording : uncontrolled	–1.982	0.666	1143.281	–2.978	<.01
uncontrolled : /p/	3.09	1.171	169.441	2.638	<.01
self-recording : /p/	4.105	0.546	1262.116	7.52	<.001

Figure 11

Interaction between segment, recording type and task type.

3.4.2 Model 5: Predicting the probability of voicing by task and recording type

For Model 5, a follow-up model was built to explore the probability of voicing across the tested conditions given the results concerning /p/. The number of observations for this model was 801.

voicing ~ recording type * task type + (1 | participant)

The results show a significant effect of recording type and task type, and a significant interaction between the two, by which the probability of voicing is greater in the uncontrolled task in self-recordings (see Table 6).

Table 6

Estimates from Model 5. The reference levels were: with experimenter for recording type and controlled for task type.

Variable	Level	Estimate	std. error	z value	p value
(Intercept)		0.629	0.242	2.598	<.01
recording type	self-recording	–0.637	0.209	–3.046	<.01
task type	uncontrolled	0.456	0.189	2.415	<.05
recording type: task type	self-recording : uncontrolled	2.428	0.655	3.709	<.001

As for the interaction, the effect of task is significant only in the self-recording condition (see Figure 12). This is confirmed by pairwise contrasts estimated based on the marginal means from the model, which show significant differences between all the conditions, except for the difference between the controlled and uncontrolled task in the with experimenter condition (β = –0.456, z = –2.415, p = .074, see Table 15 in the Appendix).

Figure 12

Interaction between recording type and task type in predicting the probability of voicing in the dialect.

3.4.3 Model 6: Predicting the probability of deletion by task and recording type

Finally, for Model 6, a follow-up model predicting the probability of deletion depending on recording type and task type was built. The number of observations for this model was 864.

deletion ~ recording type * task type + (1 | participant)

The results of the model show a significant main effect of task type and recording type (see Table 7). There was also a significant interaction between recording type and task type. This is illustrated in Figure 13, below, where we can see that task type is relevant in predicting deletion only in the with experimenter condition. The predicted probability of deletion is higher in monologues compared to sentences. The significance of this difference was also confirmed by pairwise contrasts based on the marginal means from the model, which show significant differences between the controlled and uncontrolled task in the with experimenter condition (β = –1.456, z = –7.169, p < .001), and between the controlled task in self-recordings and uncontrolled task with experimenter (β = –0.917, z = –3.787, p < .001, see Table 17 in the Appendix).

Table 7

Estimates from Model 6. The reference levels were with experimenter for recording type and controlled for task type.

Variable	level	estimate	std. error	z value	p value
(Intercept)		–1.865	0.265	–7.042	<.001
recording type	self-recording	0.539	0.258	2.092	<.05
task type	uncontrolled	1.457	0.203	7.169	<.001
recording type : task type	self-recording : uncontrolled	–1.321	0.423	–3.122	<.01

Figure 13

Interaction between recording type and task type in predicting the probability of deletion in the dialect.

4.1 Comparison between recordings and the categoricity of weakening

Both data exploration and inferential statistics show that the general intensity profiles of the investigated segments are comparable. Both /p/ and /b/ are produced with similar intensity difference values within their own categories across recordings. Furthermore, /p/ has a greater range of values, as it can be either voiced or voiceless on the surface. When we look at surface sounds, there is a clear three-category division in all tested labials. What is particularly interesting is that the two subcategories of /p/, i.e., [p] and [b], have a lot of overlap in the sentence condition. In other words, when speakers are asked to either read out scripted sentences or repeat them after another speaker, they do not seem to have two categories of segments (weakened vs. unweakened), but rather the two overlap to a great extent in terms of intensity difference despite there being voicing and lack of burst in only one of them. In spontaneous speech, however, speakers do seem to have three competing surface categories, which suggests a switch in the phonological (allophonic) category between competing forms. It must be noted, however, that fieldwork recordings explored in the second part of the analysis break this pattern. Although they contain spontaneous speech, there is an overlap between voiced and voiceless realizations of /p/ (see Figure 16 in the Appendix). This is perhaps due to the presence of the experimenter and may speak to the importance of the recording situation. The analysis of voicing reported in Section 3 showed that task type seems to be especially relevant in self-recordings, while productions made in the presence of the experimenter are more similar between tasks, which may point to the emergence of categoricity in surface forms in unsupervised spontaneous data.

4.2 Inter- and intra-speaker differences, and the observer’s paradox

The data suggest that /p/ weakening is more prevalent in spontaneous speech, although not all speakers follow this pattern (see Figure 7). This may be due to many factors, such as individual differences in speaker awareness or reactions to the speaking or recording situation. During the field experiment, I noted that while some speakers do not change their speaking patterns much when being recorded and performing the unnatural task of reading sentences, others get nervous, make more mistakes and have to repeat several of the sentences, or use a lot of hypercorrections. Also, there might be an effect of familiarity with the experimenter: Some participants may have felt less inclined to ‘perform well’ compared to others. Finally, in the self-recordings, two of the speakers were a bit nervous (voice trembling at times), and one of them (P6) seemed to have a prepared speech for the monologue, which is probably why she had an exceptionally high speaking rate in this condition.

If we accept that spontaneous speech is closer to natural productions made every day in the community, then the increased frequency of voicing in the monologue is evidence that we indeed gathered appropriate data that show more naturalistic speech closer to that ideal. It should not escape our attention, however, that there exists a discrepancy between self-recorded sentences and the self-recorded monologue in almost all speakers, even though their productions were recorded in the same session. Also, although these recordings have overall the slowest speech rates, the greatest number of /p/ weakening is produced precisely in them. This observation is also confirmed by the category distinctions illustrated on the density plots (Figure 2), where we see a discrepancy in changing surface category from [p] to [b] between the monologue and repeated sentences, and a clear distinction between spontaneous productions (monologue + IM) and sentences in general. Similar tendencies can be seen in the prevalence of /b/ deletion depending on the recording. The probability of elision is higher in self-recorded monologues, and higher still in IM. It is also higher in spontaneous speech from fieldwork interviews compared to the field experiment (see Figure 13).

These results lead to the conclusion that perhaps the experimenter (or observer) is not the most important element to remove from speech production recordings to capture natural speech. The data show that the task, or the way language is used by the speakers, plays an even more important role in influencing their speech production in the context of the observer’s paradox. Scripted sentences help control the investigated conditions and elicit exactly what we need as researchers but lead to a substantial change of style on the part of the speaker, even when sentences are repeated from memory, and even when the speaker is the one making recordings and controlling the situation.

As noted in the Introduction, recording environment can influence study participants’ speech patterns, possibly suppressing certain sound processes. However, the data presented in this paper show that, apparently, entering a ‘recording situation’, in which the participant is aware that their speech production will be subject to post-hoc evaluation, is not sufficient to explain differences in speaker productions. An analysis of fieldwork recordings of spontaneous speech demonstrates an interplay between recording type (situation) and task type. Having no flexibility over what is being produced diminishes the probability of applying a weakening process (be it /p/ voicing or /b/ deletion), although it does not necessarily affect the way particular variants are produced acoustically (intensity difference). Thus, it is possible that speech planning is organized differently depending on the type of task and input speech. Our cognitive and motor mechanisms responsible for changes in pronunciation, such as coarticulation, gestural undershoot and overlap, and other issues, may be highly dependent on whether we are merely repeating sentences written by someone else or spontaneously creating our own. This unsurprising result finds support in previous work in sociophonetics (e.g., Lewis, 2001; Hualde et al., 2011; Lozano, 2021). Repeated and redundant words or phrases are also different from the point of view of speech production (Clark & Wasow, 1998; Aylett & Turk, 2004). There has also been some work on the differences in speech planning under lexical competition or in error making conditions, focused on lab vs. spontaneous speech (e.g., Pouplier & Goldstein, 2010; Alderete et al., 2021). Different speech planning mechanisms and the resultant motor arrangements, and perhaps differences in other language-external factors modulating speech, such as social factors, attitudes, speaking habits, and so on, may result in different weakening patterns and influence researchers’ generalizations concerning sound change.

In the present case, were we to look at scripted sentences only, we might consider /p/ weakening in the dialect as much less prevalent and much more continuous or unstable than it is in natural productions. We might also underestimate the frequency of /b/ deletion. The self-recordings employed in this study helped us estimate the true scale of the studied processes and uncover additional factors influencing speaker productions. They point to the differences in the frequency of process application across different tasks and, hence, speech styles, as well as the fact that some weakening parameters remain virtually unaffected, regardless of the task or recording type. In our case, intensity difference, a marker of degree of weakening, proved not to be that sensitive to the data gathering process. More research is needed to address these questions more thoroughly, and new research paradigms should be sought to test different task types and induce more controlled or research question-aligned, yet spontaneous, productions. The use of turn-taking and speaker interactions is perhaps a good direction to follow (e.g., Lozano, 2021, successfully used conversation dyads in her study design). Interactive tools and applications could also be explored. If we want to make use of self-recordings, reading maps, giving instructions or other types of semi-structured speech could be elicited in a way similar to research on intonation. Additionally, care should be taken to disentangle the experimenter/interlocutor from a perceived interlocutor. Since we can see a difference between self-recorded monologues and instant messaging in this study, it may be that when participants know a researcher will analyze their speech recordings, they suppress their natural productions to some extent, albeit not as much during a study made with the experimenter present.

The present study offers sufficient evidence that self-recordings can be used successfully in studying some types of sound change, such as stop weakening, and that self-recorded monologues are quite close to the natural productions we might hear over an IM app, which should be elicited at least as a comparative sample. The study itself touches on some important questions concerning the viability of remote data collection (see Section 4.3, below) on the one hand, and the importance of using authentic productions, and comparing lab and spontaneous speech on the other, which is crucial for us to see whether the conclusions taken from controlled studies can be supported with more naturalistic data. The latter endeavor is not new, and the debate over lab vs. uncontrolled speech is still far from over. As Wagner et al. (2015, p. 10) put it in their introduction to a special issue on the topic in the Journal of Phonetics, “the best way to approach research questions is by embracing a certain methodological pluralism.” Depending on the phenomenon we want to study, various methodological settings can be considered. Care should also be taken to elicit as much speech as possible to obtain enough data for a comparative sample when necessary. We have seen that not many tokens of sounds of interest have been found in short monologues and, moreover, total numbers of sounds analyzed differ from speaker to speaker. Perhaps using certain topics or keywords would be a good option if we can predict that certain groups of sounds will be more frequent in a sample using a specific grammatical or semantic category.

4.3 Challenges and recommendations

Many advantages of remote data collection have been described in this paper, but researchers should also take the challenges into account and address them. One challenge is finding an adequate number of participants that can provide data according to the researcher’s instructions. In remote recordings, the load of performing the study is transferred from the researcher to the prospective participant. Instead of coming to a lab or place of encounter, the participant has to prepare the equipment and setup, as well as download software that will enable quality recording. They also have to make sure that the recording is saved in a correct format and send it to the researcher, which is sometimes complicated, as larger files cannot be easily sent by email. Thus, some participants may be initially interested in sending in their data, but never find time for it. Moreover, remote data collection cannot be performed on the go; place and time are important factors, as is ambient noise, which can also discourage prospective participants, leaving them unwilling to complete the task. Compared to an online study or survey that asks participants to simply click on the correct answer, a production study conducted in the way described here is a nuisance. In the case of the present study, several people promised to perform the task but never did it, despite being reminded for two months after the recordings were due to be shared with me.

To encourage participants to complete the study, incentives, such as remuneration, prizes, or other perks, should be considered to ensure a representative sample. Or, use third parties to assist in data collection and plan ahead to have more time before the data is in. Also, it might be useful to invite twice as many people as necessary to participate. Getting a larger sample will help address such issues as bad audio quality resulting from the participants’ non-compliance with the instructions, and other, independent factors. For example, reverberation in the room, uncontrolled background noise, and so on. Rejecting some participants from the sample will not affect statistical analysis if more recordings are made than absolutely necessary to ensure a good signal-to-noise ratio and study effect size. In this study, the sample size was small overall, especially in the IM condition. This is because segments other than /p b/ were excluded from the dataset for better data comparability. Larger sample sizes should be used in future investigations.

For some participants, following instructions correctly can pose a challenge. In our case, each participant got an email with a full package containing recorded sentences to repeat and instructions with precise information on how to proceed. Screenshots were provided illustrating where to click in Audacity or mobile app and how to save the data in the correct format. Still, three of the participants had to re-save the files and/or resend them because they were provided in an incorrect format. Thus, even though most people are familiar with the use of smartphones and computers, certain tasks will not be intuitive for them. It is therefore worth checking on the participants and repeating crucial information several times.

Another challenge is the lack of control of certain recording parameters. In our study we saw a difference in overall intensity depending on recording type. Self-made recordings were slightly louder than the ones from the field experiment, which may be due to a smaller distance of the microphone from the mouth in the former case. On the other hand, IM recordings, which are typically made with the mouth close to the smartphone, had lower intensity values compared to all other recordings. Hence, we cannot reliably say that the distance between the mouth and the device is indeed responsible for the discrepancies. In any case, researchers using remote data collection methods should proceed with caution, as we cannot be absolutely sure how sound is processed via the different IM and social media apps. Furthermore, we have no way of determining whether or not the speaker maintains a consistent distance from the microphone for the duration of the recording. Acoustic measurements such as intensity are particularly sensitive to such changes. On a positive note, the within-participant differences involving different underlying segments, voiced vs. voiceless realizations of the /p/ or stressed vs. unstressed syllables were virtually the same across the different recording types in this study. Using relative rather than absolute values of key parameters was certainly helpful here and should be the way to proceed whenever possible.

Finally, since some of the recordings analyzed in this study were made using a phone rather than a laptop with Audacity, there may be some concern about the differences in quality between the recordings. Upon careful file examination, it does not appear that there were substantial differences between the two types of devices used. I ran an additional statistical model to see if there was a difference between laptop and phone recordings. There was no statistical effect of device in the intensity measurements (F = 0.019, p = .895, see Figure 17). Any differences found in the data can be attributed to the difference between underlying segments rather than the device used. Thus, we may conclude that both laptops and smartphones are suitable for the purposes of analyzing stop lenition in Spanish.

While unguided remote data collection can pose a challenge, it is nevertheless a useful research method for phoneticians, phonologists and sociolinguists. There are obvious advantages of gathering production data this way: It is less resource-intensive, it does not require travelling arrangements and captures more data in a similar amount of time. Additionally, it eliminates researcher bias and helps us get samples of casual, everyday speech. However, as noted in Section 4.2, we should be aware of the possibility that the absence of a third party does not necessarily eliminate the problem of unnatural data. Drawing on the results of this study, I see three parts of the equation in gathering naturalistic speech productions: the setting (recording situation and speaker’s awareness of it), the observer (researcher, third party present) and the material/task used. In most studies, it is impossible to eliminate all the three obstacles to natural speech production, hence cross-comparisons of different types of recordings can help mitigate the observer’s paradox. We have seen in this comparative study that speaker behavior changes quite substantially in terms of frequency of applied weakening when they are speaking freely, without preconceived words or phrases or frame sentences. At the same time, the only recording type that does not imply an observer but does imply an audience (i.e., the recipient of the message) at the time of recording is a casual voice recording sent via an IM or social media app. By contrast, self-recordings made for a researcher have an implied observer and may still induce a change of style. Thus, crowd-sourced data consisting of social media messages might be helpful in uncovering the true range of productions and their frequency in a given speech community, while controlled recordings can be used to explore acoustic detail in a more reliable manner.