2.2.1. Terminological criteria.

Studies were considered relevant if the voice quality under investigation was labelled as any of the following: ‘creaky voice’, ‘creak’, ‘vocal fry’, ‘glottal fry’, ‘pulse(d) register’, ‘pulse(d) phonation’, ‘laryngealized voice’, ‘laryngealization’, ‘glottalized voice’ or ‘glottalization’. The justification for treating each of these terms as essentially synonymous is the pervasive, explicit reference to their synonymy throughout the literature [3,16].

It is acknowledged that a number of ‘creaky-sounding’ phonation (sub)types have been described in the literature, with some taxonomies (e.g. [17]) using some of the terms listed above to distinguish between subtypes. However, delineations of creaky voice subtypes are typically based on the observation that ‘creaky-sounding’ phonation has various physiological underpinnings and/or acoustic manifestations (see [17] for more on this). Whether different creaky voice subtypes are auditorily distinct warrants further investigation, but experimental work so far supports the treatment of creaky voice as a unitary auditory phenomenon [18]. Among sociophoneticians and SLPs, creaky voice is overwhelmingly conceptualised as a singular percept.

2.2.2. Sample criteria.

Papers were included if their speaker sample and sampling methods met the following inclusion criteria:

1. The speakers must have been ‘native’ speakers of English (i.e. they were not described as ‘English learners’ or ‘L2 English speakers’);
2. The speakers were not purposefully selected for participation based on their use of creaky voice (as they were in [9], for example);
3. The speakers could be assumed to be vocally-healthy (i.e. they were not described as being in poor health or as having vocal pathology);
4. The speech analysed was English (Gibson, Summers and Walls’ [19] study of ‘English-like’ non-words did not satisfy this criterion); and
5. The researchers did not ‘interfere’ with the speakers’ performance by giving instructions regarding their use of creaky voice.

2.2.3. Analysis/reporting criteria.

Recall that creaky voice prevalence is defined here as how much phonation is realised as creaky voice. Accordingly, inclusion was limited to studies that analysed whole stretches of continuous speech and coded intervals within as either creaky voice (‘+Creak’) or not creaky voice (‘–Creak’). Note, however, that this is not the only approach that has been taken to quantifying the ‘creakiness’ of speakers’ voices. One alternative approach is to quantify the ‘creakiness’ of speakers’ voices using ordinal perceptual scale (e.g. as [20] did for speakers of Standard Southern British English); another is to extract acoustic measures like spectral tilt, from which the degree of perceptual ‘creakiness’ (and ‘breathiness’) can be inferred (e.g. as [21] did for speakers of New Zealand English). Such studies are not included in this review because they do not offer an answer to the question what is the prevalence of creaky voice in English? Instead, they answer a similar—equally valid, but fundamentally different—question: how creaky do English speakers’ voices sound?

The following types of studies were also not included: qualitative accounts of creaky voice use (e.g. [22]); studies that reported the statistical relationships between the occurrence of creaky voice and variables of interest (e.g. sentence length) rather than actual prevalence data (e.g. [23]); and studies reporting prevalence of creaky phonation but only in specific phonemic/prosodic contexts or ‘hand-picked’ words (e.g. [24–26]).

2.2.4. Publication criteria.

Finally, inclusion was limited to reports of original research distributed by peer-reviewed academic publications (journals or peer-reviewed conference papers). The purpose of limiting eligibility in this way was to establish a consistent sampling method to identify literature likely to be of a high methodological standard [15]. Reviews, ‘viewpoint’ articles, dissertations, and other research in the ‘grey’ literature were thus not considered eligible. Papers published in languages other than English were not included as their evaluation was beyond the authors’ abilities.

3.2.1. Speakers.

Table 3 shows a summary of the number of the speakers that participated in each study, and their reported characteristics. Descriptions of the speakers were largely limited to the ‘macrosociological’ [39] characteristics of sex, age, and variety of English. With regards to speaker sex, all of the studies used a binary classification, and none made a distinction between biological sex and self-described gender. Accordingly, the terms ‘females’ and ‘women’, and the terms ‘males’ and ‘men’, are used interchangeably throughout this paper.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Summary of the samples of the ten included studies.

https://doi.org/10.1371/journal.pone.0229960.t003

All but one study were analyses of then present-day American English speech (and, as Table 2 shows, all of these studies were conducted this decade). Of all the speakers sampled (n = 297), most were women (64%; n = 189), half were American women (50%; n = 149), and of these American women, at least 55% (n = 82; [31,32,37,38]) were college students aged under 30. Thus, female speech has been studied more than male speech, with a ratio of approximately 3:1, and young adults are more studied than older adults. The most studied English speakers in terms of creaky voice prevalence are young, female, American, college students.

Many of the American English studies described their participants as speakers of a particular variety or dialect of American English (see the last column in Table 3). Of these, Yuasa [30] provided the most detailed description of the sampled speakers’ demographics: all had resided in California for at least five years, but not all were born and raised there, and the group was mixed in terms of speakers’ ethnicities (see Table 1 of [30]). However, in the other American studies, detail regarding the speakers’ residential histories and ethnicities was minimal [32,37,38] or absent [31,33–36].

Henton and Bladon’s [29] study was the only identified study to describe creaky voice prevalence in a non-American variety of English. The study analysed speech recordings from a corpus of British English speakers recorded in the mid-1980s [40] in which speakers were grouped into two accent groups: Received Pronunciation (RP), and Modified Northern (MN). As suggested by Henton and Bladon [29], more information about the speakers and the criteria used to group speakers into the two groups can be retrieved from [40 pp172-175]; in short, the speakers were described as ‘white’ and ‘academically employed’, and grouped based on a perceptual judgement of accent, as well as residential history (the MN speakers had ‘modified’ their accents because after being raised in the area of Leeds, Yorkshire, they had moved to the South of England).

3.2.2. Speech material.

Six studies [29,31,33–35,38] analysed read speech only, two studies [30,36] analysed spontaneous speech only, and two studies [32,37] analysed both read and spontaneous speech. Among the eight studies that analysed read speech, the nature and length of the text that participants were asked to read aloud varied. Henton and Bladon’s [29] speakers read aloud eleven stand-alone sentences. Luthern and Clopper’s [34] speakers read aloud five passages, of 35–90 seconds in duration, twice—once in a ‘plain lab speaking style’ and once in a ‘clear lab speaking style’. The names of the passages were not reported in the paper. Melvin and Clopper’s [35] speakers also read aloud five unspecified passages, each approximately 60 seconds in duration, but only once. Benoist-Lucy and Pillot-Loiseau’s [32] speakers read aloud The North Wind and the Sun passage [41], three times. The remaining four studies that analysed read speech [31,33,37,38] had their speakers read The Rainbow Passage [42], but varied in how much of passage they analysed. Borrie and Delfino [37] presumably analysed its entirety, though whether this means just the first six sentences of the passage, as is commonly done in phonetic research, or the whole passage [42], is not specified. The two studies by Abdelli-Beruh and colleagues [31,33] analysed the 2^nd, 4^th, and 6^th sentences of the passage. Cantor-Cutiva, Bottalico and Hunter [38], whose participants read the passage aloud in nine different simulated noise conditions, analysed only its 6^th sentence in each condition.

Among the four studies that analysed spontaneous speech [30,32,36,37], there was also variation in the nature and duration of the speech material analysed. Yuasa [30] recorded her participants having an approximately 10-minute conversation on the topic of food, but analysed only a randomly selected 401-word segment from each conversation. Benoist-Lucy and Pillot-Loiseau [32] analysed 15-minute conversations on a topic of the participants’ choosing; whether the two interlocutors spoke for equal portions of the 15-minutes was not reported. Oliveira et al. [36] analysed monologues of unspecified duration in which participants described all the steps involved in making a peanut butter and jelly sandwich and in doing laundry. Borrie and Delfino [37] analysed speech produced during the first 5 minutes of a cooperative picture-matching task; how much speech was produced by the participants during that 5 minutes (as opposed to that produced by the interlocutor) was not reported.

In the studies that analysed spontaneous, conversational speech (i.e. not a monologue) [30,32,37], there were various controls placed on the conversational partner. In Yuasa’s [30] study, some but not all speakers knew their conversation partner, and all of the women conversed with speakers of the same sex, whereas eight of the eleven men conversed with speakers of the opposite sex. In Benoist-Lucy and Pillot-Loiseau’s [32] study, the (all-female) participants conversed with same-sex interlocutors, all of whom were also participants, and with whom they were familiar. In Borrie and Delfino’s [37] study, the (all-female) participants spoke with two different interlocutors: one woman who had been chosen for her tendency to produce a ‘substantial’ amount of creaky voice while talking, and another woman for her tendency to use a ‘minimal’ amount. Thus, for Borrie and Delfino’s [37] study—a study of ‘conversational entrainment’—the amount of creaky voice in the conversational partner’s speech was a variable of interest.

3.3.1. Methods for detecting creaky voice.

Three types of ‘detection methods’ (that is, methods used to decide whether intervals of speech should be coded as +Creak or–Creak) have been used in creaky voice prevalence studies:

1. Auditory methods (used in [29,31,33,37,38]): one or more person listened to the audio recordings and identified the occurrence of creaky voice using pre-determined auditory criteria.
2. Auditory-visual methods (used in [30,32,35,36]): like auditory methods, but in addition, at some point during analysis, the rater/s visually examined one or more type of acoustic/physiological analysis (e.g. a spectrogram) to assist in coding decisions.
3. Automated acoustic methods (used in [34]): an objective, acoustic criterion was used to code speech as +Creak or–Creak.

In the studies that used auditory or auditory-visual methods, the auditory criteria for what constituted +Creak included: a perceivably low pitch [31,33,35,37]; the perception of a ‘rough’ or ‘gravel-like’ quality [31–33,37]; a train of separately resolvable glottal pulses [29,30,32]; and/or the impression of a ‘popping corn’ sound [30,38]. One study [36] defined creaky voice as “a creaky sound with a rough vocal quality that is produced at a low fundamental frequency” in the paper’s introduction, but did not specify precisely what auditory criteria the raters used during their analysis.

Among the studies that used auditory-visual methods, two [32,36] did not explicitly state what constituted visual evidence of creaky voice. However, two did: in those studies, visual evidence of creaky voice was increased irregularity of pulses in the speech waveforms [30,35] and/or widely-spaced vertical striations in a spectrogram representation [30].

The sole study that used an automated acoustic method [34] detected creaky voice using the following procedure: after using a forced aligner with manual correction to identify each syllable nucleus, ten equally-spaced fundamental frequency (F0) measures were extracted from each vowel. A syllable was coded as a +Creak syllable if the majority of the F0 measures were <150Hz. The justification for this threshold was that it was “based on previous manual coding of voice quality in these data” [34].

3.3.2. Formulae for calculating creaky voice prevalence.

The ten included studies calculated creaky voice prevalence using five different prevalence formulae: (note that [32] used two formulae)

1. (used in [31,33,38])
2. (used in [30])
3. (used in [29,34,35])
4. (used in [32,37])
5. (used in [32,36])

In the seven studies that used formulae 1–3 [29–31,33–35,38], the recordings of continuous speech were first partitioned into units: sentences, words, or syllables. Each unit was then coded as either +Creak or–Creak according to the study’s method for detecting creaky voice (see Section 3.3.1.), and prevalence then calculated as the percentage of units coded as +Creak.

In the studies that used formula 4 and/or formula 5 [32,36,37], the whole speech recording (or that portion of the recording selected for analysis) was examined in Praat [43] as a continuous signal, and the boundaries of each interval of creaky voice were annotated in a Textgrid file, according to the study’s method for detecting creaky voice (again, see Section 3.3.1.). For formula 4, the durations (in ms) of all annotated +Creak intervals were summed, and prevalence calculated in percentage terms as the total duration of +Creak in relation to the ‘total duration of speech’. For formula 5, the number of intervals annotated as +Creak were counted, and prevalence calculated as the number of +Creak intervals per minute of the speech signal. It was not reported precisely how ‘total duration of speech’ or ‘minutes of speech’ (the respective denominators for formulae 4 and 5) were calculated. Both require a measure of total time spent speaking, but it was not made clear if, for example, momentary pauses between words and phrases were deducted from the calculation.

3.4.1. Variation across groups.

Three studies contribute some insight into how creaky voice prevalence patterns across speakers grouped by age. Oliveira et al. [36] reported the number of creaky voice occurrences per minute in the monologues of 40 American women across two age groups: younger women aged 18–25 (n = 20), and older women aged 35–50 (n = 20). The study found a slightly but not significantly higher rate of +Creak occurrences in the speech of the younger women (Fig 2; columns 21–22). Benoist-Lucy and Pillot-Loiseau [32] also reported the number of creaky voice occurrences per minute (also using an auditory-visual method) in their sample of 20-year-old American women (n = 6). In comparison to Oliveira et al.’s [36] older group, Benoist-Lucy and Pillot-Loiseau’s [32] 20-year-old speakers produced creaky voice occurrences at a considerably higher rate when conversing with a friend (compare Fig 2; columns 20,22). But see Fig 2, columns 19 and 22—when Benoist-Lucy and Pillot-Loiseau’s [32] speakers were reading aloud, they produced creaky voice occurrences at a similar rate to Oliveira et al.’s [36] older group. Cantor-Cutiva, Bottalico and Hunter [38] conducted multivariate statistical analysis to see whether age was associated with creaky voice use, and found no association. However, this study only sampled speakers aged 20 to 25, arguably a somewhat narrow range for testing age-based sociophonetic differences.

Comparatively more studies have investigated how creaky voice prevalence patterns across speakers when grouped by sex. The studies on American English speakers [30,31,33,35] found that female speakers on average produced more +Creak sentences (Fig 2; columns 1–2), more +Creak words (Fig 2; columns 4–5), and more +Creak syllables (Fig 2; columns 11–12) than male speakers. However, the sole study on British English speakers [29] found an opposite pattern (Fig 2; columns 6–9), and particularly so in the Modified Northern (MN) accent group, in which men produced +Creak syllables at a rate approximately 6 times higher than their female counterparts (Fig 2; columns 8–9).

It is tempting to conclude, then, that the prevalence of creaky voice patterns across sex patterns in opposite directions in American English versus British English. However, in comparing the American English studies with the British English study, there are numerous confounding variables. First, the American English studies—those that reveal something about prevalence across sex groups—were conducted in 2010 [30], 2012 [31], 2014 [33] and 2015 [35], and were analyses of then present-day speech recordings; the British English study [29] analysed recordings collected in the mid-1980s. Second, the age ranges of the samples are not the same: the groups of American speakers were aged 18–25 [31,33,35] and 18–34 [30]; the British English speakers were aged 25–40 [29]. A third confounding variable is the range of speech tasks used across the studies (see Section 3.2.2).

3.4.2. Variation across individuals.

Five studies reported results in ways that provide insight into the extent of variability across individuals. Two studies [32,34] reported the prevalence scores of individual speakers—in both, the scores reported were pooled scores for two different speaking tasks: Benoist-Lucy and Pillot-Loiseau [32] reported (on a graph, without precise figures) individual scores for read speech and conversational speech combined (Fig 2; column 16); and Luthern and Clopper [34] reported individual scores for two reading styles (‘plain’ and ‘clear’) combined (Fig 2; column 10). Three additional studies [35–37] revealed there was within-group variance by reporting standard deviation (SD) values (Fig 2; columns 11–12, 14–15, and 21–22). Taken together, these five studies show that there is variability across individual speakers in the prevalence of creaky voice, even within relatively homogeneous socio-demographic groupings. However, individual variation was not their focus, and they reveal nothing about which individuals deviated far from the group’s mean. Cantor-Cutiva, Bottalico and Hunter’s [38] statistical analysis investigated the association between creaky voice use and a range of individual ‘lifestyle’ and ‘health’ factors, and found some association between speakers’ habitual consumption of caffeine and the occurrence of creaky voice during the final sentence of a reading passage.

3.4.3. Intra-speaker variation, across situations/contexts.

Three studies [32,37,38] measured the prevalence of creaky voice across more than one task or speaking context. Benoist-Lucy and Pillot-Loiseau [32] found that, in their sample of 20-year-old American women (n = 6), the average percent of speech produced as creaky voice during conversation was more than double than that while reading aloud (Fig 2; columns 17–18). Borrie and Delfino [37] also found creaky voice to be more prevalent in the conversational speech of young American women (aged 18–29, n = 20) compared to their read speech, but only when their interlocutor was someone who spoke with a ‘substantial’ amount of creaky voice themselves (Fig 2; columns 13–15). (Statistical analysis showed that the extent to which the participants socially aligned with the interlocutor was also a contributing factor—see [37] for more on this). Finally, Cantor-Cutiva, Bottalico and Hunter’s [38] experimental study found that background noise conditions may affect speakers’ likelihood of producing creaky voice (Fig 2; column 3). Together, these three studies suggest that the prevalence of creaky voice tends to be higher in spontaneous speech, but moderating factors include who the interlocutor is, the speakers’ relationship with the interlocutor, and the acoustic conditions in which the conversation is taking place.

4.1.1. Choosing the sample.

This review found that college-aged American women are the most studied group of English speakers in terms of creaky voice prevalence. This imbalanced focus, which has only amplified over the last decade, seem to be partially driven by beliefs surrounding a recent increase in prevalence in this group; in one of the prevalence studies reviewed herein, the authors explained their decision to sample only American female college students by saying that this “corresponded to the literature consensus stating that the creaky voice effect is more robust among American young women and often occurs in a College context” [32 p2396]. However, this review found that such a consensus is not substantiated with quantitative evidence. In fact, the relationship between beliefs surrounding creaky voice and the design of research seems to be circular; as Lawson [45] has commented, researchers’ focus on creaky voice among young American women has fuelled public discourse on the topic.

Investigations of less studied populations should therefore be prioritised. This includes studies of older speakers, male (or non-binary) speakers, speakers of other English varieties and, if possible, studies of historical speech recordings, too. Because the research also shows that creaky voice prevalence can vary within speakers, across situations and tasks [32,37,38], it is vital that future researchers control for and describe in detail the nature of the speech sample under analysis.

4.1.2. Choosing a detection method.

This review found that most prevalence studies have used auditory or auditory-visual methods to detect occurrences of creaky voice. This is quite justified, given that creaky voice is chiefly a perceptual phenomenon. However, using such methods makes analysis of long speech recordings and/or across a large number of speakers time- and labour- intensive. (In this light, it makes sense that the studies identified by this review were all small in scale). The studies that used auditory or auditory-visual methods and reported levels of inter- and intra- rater agreement (either formally or informally) reported that agreement was often high, but not complete [29,31,33,36–38]. This is in line with evidence more generally of differences across and within listeners in the perceptual judgement of voice quality [20,46].

The use of automated acoustic methods for detecting creaky voice has potential to enable more replicable, larger-scale studies [5,47], but the acoustic complexity of creaky voice [17] means that the reliability of such methods needs close examination. This review found that only one creaky voice prevalence study [34] has made use of such a method. The method used was based on the use of a F0 cut-off point or ‘threshold’ (150Hz) to differentiate between +Creak and–Creak. Since most manifestations of creaky voice are characterised by low F0 [17], it is conceivable that phonation can be classified as being either +Creak or–Creak, with a high degree of approximation, using an automated F0-based method [48].

4.1.3. Choosing a prevalence formula.

Most studies identified in this review calculated prevalence of creaky voice as the percentage of units of analysis (sentences, words, or syllables) coded as +Creak. A ‘unit’-based formula is typical for quantifying the prevalence of allophonic variants of specific phonemes [7]. However, creaky voice is a suprasegmental feature of English [6], so its manifestations do not necessarily align with the boundaries of sentences, words, or syllables. If a researcher does choose to calculate creaky voice prevalence as a percentage of +Creak units, they must face the question: which type of unit (sentences, words, or syllables–or indeed something else) is best? One thing to consider is that the ‘size’ (duration) of a unit is inversely proportional to the amount of insight it provides. For example, dissecting continuous speech into large units like whole sentences ‘pixelates’ the data, and gives no insight into how much of each ‘pixel’ is realised as creaky voice; smaller units, such as syllables, give a comparatively higher resolution.

Two other formulae for calculating prevalence were found in the literature. One was the number of +Creak occurrences per minutes of speech. This formula arguably offers the least insight in terms of the actual amount that creaky voice occurs during speech; hypothetically, the same amount of creaky voice might manifest as one single occurrence, or several separate occurrences of shorter duration.

The final formula found in the literature was the percentage of speaking time realised as +Creak. A strength of this formula is that it would be comparatively less impacted by very short durations of creaky voice associated with glottalised realisations of specific phonemes [24,49]; as for the other formulae, such momentary glottalisations might cause whole syllables, words, or sentences to be coded as +Creak. Or they might be counted as whole +Creak occurrences. Whether glottalised realisations of specific phonemes ought to ‘count towards’ creaky voice prevalence measures is worth debating. No quantitative creaky voice prevalence study identified by this review explicitly differentiated between very momentary manifestations of creaky voice associated with specific phonemic segments, and suprasegmental manifestations known as ‘phrasal creak’ [50].

4.1.4. Reporting the results.

Most studies identified in this review did not report creaky voice prevalence scores for individual speakers. Instead, only means were reported for the whole sample or for sub-groups within the sample. In many cases, not even measures of variance around the group means, such as SD values, were reported. However, studies that did report the findings in a way that revealed within-group variation [32,34–37] showed that creaky voice prevalence varies across individuals, even within relatively narrow demographic groups. Future studies that aim to be optimally informative should therefore report within-group variance whenever possible. Doing so may enable a future meta-analysis.