The role of place and manner of articulation in Kurtöp tonogenesis: refining the model

1.1 Tone and tonogenesis

Yip (2002: 1) distinguishes tone from intonation when ‘the pitch of the word can change the meaning of the word’… in terms of its ‘core meaning’. Traditional definitions of tone are often associated with individual syllables, though many languages only have tone at the word level. DeLancey (2003), for example, describes this type of system in Lhasa Tibetan.^[1] This type of melodic tone system is seen in many tone languages around the world, including in Kurtöp (discussed below).

Linguistics has long been concerned with how sounds, including tones, change over time, dating back at least to the Neogrammarian hypothesis of the late 19th century. As Phonetics advanced as a discipline, it became increasingly apparent that phonological contrasts are often realised and perceived by a suite of different acoustic features, making the issue of sound change more complex and harder to quantify. For example, ‘voicing’ in English can be reliably realised by VOT in word-initial position but in word-final position the contrast is often saliently seen in the length of the preceding vowel. Duration of stop closure is also a relevant parameter that can be used to capture the distinction between ‘voiced’ and ‘voiceless’ English stops. Multiple acoustic cues have been particularly noted in Korean lenis stops (e.g. Silva 2006; Kang and Guion 2008, and many follow up studies) and other languages such as Cao Bằng Tai (Pittayaporn and Kirby 2017), Assamese (Dutta and Kenstowicz 2018) and Dzongkha (Kirby and Hyslop 2019), amongst many others. The issues of these multiple cues and how they are utilised in speech can be referred to as ‘cue-weighting’, which is succinctly summarised by Schertz and Clare (2019). They consider cue-weighting in both production and perception and discuss the theoretical link between the two, and along the way provide several clear illustrations of cue-weighting, or how speakers and listeners produce and perceive the most important cues along many competing dimensions.

The issue of multiple cues is also relevant for studies of sound change. It is established that in contexts where multiple cues are documented to be part of a phonemic contrast, the issue of which cue(s) are most salient in production or perception is non-trivial. This applies especially along the diachronic dimension, where one cue may be found to be replacing another as having primary importance. Kirby (2013) examines situations in which multiple cues are involved in sound change. He takes the example of Seoul Korean, in which VOT, f0, spectral tilt and amplitude of the release burst are all potentially relevant perceptual cues to denote the laryngeal setting of the contrast in onset position (Cho et al. 2002; Wright 2007), yet only f0 appears to be the relevant contrast for young speakers today. We can consider such cases as examples of transphonologisation, in which the reliance has shifted from one cue to another over time. The term was first coined by Hagège and Haudricourt (1978), to describe the replacement of one distinctive opposition with another. Looking at sound change in a complex phonetic domain, we can consider transphonologisation to be the shift from primary weighting of one cue to that of another. We further make the assumption in this article – describing a case of tonogenesis – that the shift from a primary acoustic cue of voicing (VOT and presence of voicing in fricative onsets) to a primary cue of f0 on the following vowel represents an example of transphonologisation, signalling a phonological shift, from that of the consonantal domain (voicing) to that of the suprasegmental domain (tone).

Tonogenesis, or the birth of tone, has already been discussed in many languages, including Athabaskan languages (Kingston 2005), Chinese (Chen 2000), Cèmuhî (Rivierre 1993), Kammu (Svantesson and House 2006), Lahu (Matisoff 1970), and Kurtöp (Hyslop 2009), amongst many others. Although the term ‘tonogenesis’ was first used by Matisoff (1970) to describe the phenomenon in Lahu, the discussion of tone from a diachronic perspective probably began with Maspero (1912) and was brought into the modern linguistic spotlight by the prominent work of Haudricourt (1954)’s analysis of Vietnamese tone. Using Vietnamese, Haudricourt discusses tone development through a series of sound changes. These steps, listed below, have become the primary model of tonogenesis and are found widely in the literature on tonogenesis.

1. Contour melodies develop on syllables ending in laryngeal consonants, [h] and [ʔ], while syllables without laryngeals either maintain a level pitch, or in the case of syllables with a stop final, did not exhibit any tonal contours

2. The laryngeal finals are lost over time and three tonal environments phonologise in their absence: rising (ʔ), falling (h), and level (non laryngeal, non stop finals)

3. Tone splits result from the merger of voiceless and voiced initials and associated laryngeal features on the following vowel

(Haudricourt 1954)^[2]

Parts of this model can be found in many examples of tonogenesis. The first two steps involving lost codas are less prevalent around the world, while the process involving initials (step 3) is the most widely reported and represented within tonogenetic literature (Ratliff 2015). The focus in this article is the development of high versus low tone registers from a voicing contrast in onsets (Haudricourt’s step 3), although this really is part of the initial introduction of tone into Kurtöp because step 1 has been specifically used for Haudricourt (1954).

The above model has been updated to account for more intermediate stages. For example, Thurgood (2002) argues that features such as ‘breathy voice’ and ‘creaky voice’ also played a role in the development of Vietnamese tone. However, this does not explain the development of tone in hundreds of other languages. In some languages, such as Arapaho (Algonquian language of central North America; Goddard 1991) and Central Tibetan (DeLancey 1989) an entire syllable was lost in words, leading to a tone on the preceding syllable. In other languages, such as Southeastern Monguor (Mongolic language of China; Dwyer 2008) and Balsas Nahuatl (an Uto-Aztecan language of Mexico; Guion et al. 2009) the stress system interacted with other changes in the language, leading to the development of tone. Other factors that have been reported to be involved include vowel length, vowel height, aspiration of consonants before and after vowels, and several others. However, very few of these processes have been studied in any depth; most of our understanding of tonogenesis in these contexts comes from only a handful of words compared across similar languages or dialects. Further, a brief survey of the literature reveals these small cases actually comprise the majority of instances of tonogenesis (Hyslop 2023). It is therefore paramount to study these cases in depth in order to advance our understanding of how and why languages develop tone.

There has also been considerable work in the past decade or so, documenting what happens with tone once it enters language, also called tone change (Dockum 2019; Ratliff 2015). The current study is interested in tonogenesis as a whole, and while tone is already established in Kurtöp, the term ‘tone change’ is less applicable. The tones in Kurtöp are not changing; rather, tone is synchronically present in a small phonological domain (following sonorant consonants and the palatal fricative) and is currently spreading to other phonological domains (following all consonants).

1.2 Kurtöp

Kurtöp is an East Bodish (<Tibeto-Burman) language indigenous to Bhutan. The East Bodish languages are closely related to Tibetic languages but are not direct descendants of Old Tibetan. The precise relationship between Tibetic and East Bodish languages is a matter of ongoing research, which is further complicated by long term, heavy contact between East Bodish and Tibetic languages (see e.g. Hyslop 2021b). There are approximately 4,000–5,000 speakers of Kurtöp^[3] in the Lhüntsi district, spread across perhaps a dozen different villages^[4] and six mutually intelligible dialects.^[5] One grammar of the language has been written (Hyslop 2017). The Kurtöp-speaking region of Bhutan is shown in Figure 1.

Figure 1:

Kurtöp-speaking region in Bhutan.

All East Bodish languages are tonal, and the current state-of-the-art shows that they all have tone systems similar to that described for Kurtöp. That is, they all have tone fully contrastive following sonorant onsets and the palatal fricative and incipient tone following the other obstruents. Kurtöp phonology was previously described by Michailovsky and Mazaudon (1994) and Hyslop (2009, 2017. This section will provide a brief summary, based on Hyslop (2017).

The nine syllable structures of Kurtöp are given in Table 1. Note that the simple, short vowel appears to be found rarely and only as one syllable in a multisyllabic word, as in the first syllable of ‘food’.

Syllables can be made up of a minimum of a rime, which can be a short or long vowel, diphthong, or a short vowel plus coda [-p, -t, -k, (-s), -m, -n, -ŋ, -r, (-l)]. The -s coda is not found in all dialects of Kurtöp and is environmentally conditioned in the varieties for which it is found. The -l coda has been lost except when conditioned by verbal suffix apocope and loan words (see Hyslop 2017). The maximum syllable shape is complex onset plus rime.

Previous work has established that Kurtöp has a three-way voicing contrast in stops between voiceless, voiceless aspirated, and voiced,^[6] and a two-way contrast in dental fricatives (voiced and voiceless). Any of Kurtöp’s 31 consonants, listed in Table 2, can be used in onset position, and an additional thirteen onset clusters are permissible: /pr-, pʰr- pj-, pʰj-, pl-, br-, bj-, bl-, kw-, kʰw-, gw-, mr-, mj-/. The onset clusters are undergoing a process of simplification, including the merger of /bl-/ and /br-/, the simplification of /py-/ to [pç∼pc∼c], of /mj-/ to /ɲ/, and of /kʰr, kr, gr/ to /ʈ, ʈʰ, ɖ/.

Kurtöp has a basic vowel system with the five cardinal vowels, /i, e, ɐ, o, u/, and four diphthongs, /ɐu, iu, ui, oi/.^[7] Michailovsky and Mazaudon (1994) included /ai/ in the diphthong inventory, but Hyslop (2017) reports that this has since undergone a sound change /ai/>/e/, in Kurtöp. It appears that Kurtöp diphthongs, like its complex consonant cluster onsets, are undergoing a process of simplification.

Kurtöp stress is predictably found on the initial syllable of a root and realised primarily through the acoustic correlate of duration.^[8] Complex onsets and long vowels are also only found on stressed syllables.^[9] Stressed syllables almost always bear tone as well, but tone can be separated from stress in that it must be word initial, while stress is root initial. The only instance in which these two circumstances are not identical is when the negative prefix (the only prefix found in the language) is used, causing the tone to move left to the word-initial syllable.

Tone in Kurtöp is found in monosyllables and in word-initial position. In the case of multisyllabic words, then, only the first syllable of the word has tone and the remaining syllables follow a predictable melody: HL or LH (only disyllables have been quantified).

Michailovsky and Mazaudon (1994) presented the first evidence for tonal contrasts in Kurtöp, showing tone as contrastive following the sonorants. They also noted the predictable tone register of the obstruent series, stating that ‘two tonal registers, high and low, correlated with voicing oppositions in the initial consonant, and remnants of initial clusters’ (546). Hyslop (2009, 2017 has corroborated, and expanded upon, this early description of Kurtöp tone.

Tone is contrastive following all sonorants and the palatal fricative, as shown in Table 3. Elsewhere, tone is predictable, with high tone following the remainder of the voiceless obstruents and low tone following the voiced obstruents; see Table 4.

The following discussion is a summary of Hyslop (2009). Kurtöp can be shown to be developing tone in the following order: 1) tone phonologises following sonorant onsets, conditioned by diachronic onset clusters; 2) tone phonologises following palatal fricatives; 3) tone now predictable following all obstruents (low pitch following voiced; high pitched following voiceless) and stops are in the process of devoicing. That is, tone has replaced a contrast in voicing following sonorant onsets and the palatal fricative; and tone is currently in the process of replacing a voicing contrast following the remainder of the obstruents.

Using comparative evidence, Hyslop (2009) demonstrated how tone first phonologised following sonorants from s- initial sonorant clusters. In Written Tibetan, a cousin language to Kurtöp, s- sonorant clusters are found in words in which Kurtöp has high tone. To these, we can add k- initial sonorant clusters as well, based on ongoing comparative work. For example, Hyslop (2014) reconstructs *kwa ‘tooth’ and *kram ‘otter’ to Proto East Bodish, which lead to wá and rám in Kurtöp, respectively. Comparative data showing the proposed consonantal sources for the high tone following sonorants in Kurtöp is shown in Table 5 below.

Hyslop (2009) speculated that the sonorant would assimilate to the voiceless s- creating a voiceless sonorant and at some point, the pitch conditioned by either the voiceless sonorant or the preceding consonant would have phonologised into tone (i.e. a transphonologisation of voicing in onset to tone on vowel). Low tone is then phonologised following the voiced sonorants. Minimal evidence has been found for this intermediate devoicing step in the Upper Mangdep (Phobjip dialect) form for ‘hair’, which is sometimes realised with a voiceless rhotic. This fits with what is known about voiceless sonorants being ‘pitch raisers’ (L-Thongkum 1997: 1082). We hope more comparative data will become available – especially for acoustic analysis – soon.^{[10] , [11]} Until then, the precise acoustic details that explain the transfer of a contrast in the consonantal onset domain to one of tone in the rime remain a matter of speculation. Nonetheless the fact remains that tone has phonologised following the sonorant consonant initials first.

The palatal fricative /ç/ is argued to have recently phonologised tone on the following vowel in Kurtöp. Michailovsky and Mazaudon (1994) remark the predictable tone register found on obstruents with the following caveat, ‘voicing is often absent in pronunciation, leaving only the low tone to insure the contrast. Thus, ʑ- is usually pronounced ᶫɕ-’ (547). The fact that Hyslop (2009, 2017 never encountered the voiced palatal fricative, only encountering the voiceless with low tone series, suggests that the palatal fricative was likely undergoing a process of tonogenesis when Michailovsky and Mazaudon conducted their fieldwork in the 1970s, and has since phonologised following voiceless palatals.

Hyslop (2009) offered acoustic evidence that tone is phonologising following the stops via an acoustic study involving two speakers. She showed that f0 on the midpoint following voiced versus voiceless stops was statistically significant for both speakers. At the same time, she also showed that the VOT values of the phonologically voiced series were merging with the VOT values for the voiceless series. Taken together, Hyslop (2009) proposes that tone first phonologised following the voiceless and voiced category of stops and that the voiced stops are now merging with the voiceless stops. Finally, she proposes a follow up study looking into the role of sonority in tonogenesis, noting that tone had reportedly entered Tibetan, Tshangla and Tai following the sonorants in a similar way to Kurtöp (Hyslop 2009: 843).

As mentioned in §1.2, previous studies in tonogenesis have found that breathy voice can be a crucial, mediating step between a contrast in voiced obstruents and low tone on following vowels. This was made explicit in Thurgood (2002) and many studies have since observed this (see for example Pittayaporn and Kirby 2017; Kirby and Hyslop 2019). However, in fifteen years of working with multiple speakers of the language, we have never once heard breathy voice uttered by a Kurtöp speaker, even those who are fluent in Dzongkha. This is in contrast to Dzongkha, the national language of Bhutan and with which the first author is also familiar. As such, no measurements have been made to quantify (its absence).^[12] In this way, Kurtöp is similar to Afrikaans (Coetzee et al. 2018), Khmu (Svantesson and House 2006) and Malagasy (Howe 2017), in showing that indeed some languages can move directly from voicing to tone without requiring mediating voice quality.

While Hyslop (2009) was able to convincingly argue that a contrast in VOT of stops was being partially replaced by a contrast in f0 on the following vowel, she was not able to investigate whether all stops behaved equally, or how the fricatives participated in the sound change.

2.1 Speakers

Two speakers were recorded in Thimphu, Bhutan, in 2009.^[13] KT is a male in his forties. He is from Tabi and in addition to being a native speaker of Kurtöp, he also speaks Dzongkha, Hindi and some English. Ch is a female in her early sixties. She is from Dungkhar, and in addition to being a native speaker of Kurtöp, she also speaks Dzongkha, Tshangla, Chocangaca, and Bumthap.

We recorded three additional speakers in Canberra, Australia in December of 2015. KL is a male speaker in his early thirties. He is from Tabi but left the village when he was six years old, though he returns annually. Since leaving, he has lived in Dzongkha-speaking Thimphu, with the exception of three years in Trashigang, two years in Paro, and two years in Canberra, Australia. In addition to being a native speaker of Kurtöp, he also speaks Dzongkha, Nepali, Tshangla, Hindi, and English. He has a master’s degree that he obtained in English instruction in Australia. DM is a female in her early thirties. She is from Shawa but left the village when she was five and has only returned three times since. She lived in Trashigang for six years, Samtse for four years, Tsirang for three years, Paro for five years, Wangdi Phodrang for seven years, and Canberra, Australia for two years. In addition to being a native speaker of Kurtöp, she also speaks Dzongkha, Nepali, Hindi, and English. SD is a male speaker in his mid-thirties. He is from Dungkhar but left when he was between fifteen and sixteen years old. He has lived in Mongar, Thimphu, Samdrupjongkhar, and Canberra, Australia. In addition to being a native speaker of Kurtöp, he also speaks Dzongkha, Nepali, Tshangla, and English. The speakers were found on a voluntary basis through personal networks.

2.2 Materials

A total of 3,540 monosyllabic consonant initial tokens were recorded and analysed acoustically. The elicitation was done using a wordlist of 216 words (see Appendix 1), elicited in English and an orthographic representation of Kurtöp. The wordlist was given to each speaker and reviewed for clarification and meaning prior to recording. The speakers then produced each target word three times and a fourth time in the carrier phrase shown in (1).

List intonation was found to be prevalent in the three words spoken in isolation prior to the carrier phrase. There was often a rising contour in the first token and a falling contour in the third token; however, similar to Hyslop (2009), since this was true regardless of the initial consonant’s manner, voice type, or place of articulation, and the analysis was based on level of the pitch and not contour, the inclusion of all tokens, despite list intonation, should affect all data in a predictable and consistent way and will therefore not skew any results.

All utterances were included in the analysis with the exception of a few omitted for mispronunciation. Occasionally the target word was repeated four or five times prior to the carrier phrase, which yielded additional tokens for the word within that speaker. Sometimes, a speaker would not be familiar with a target word, and it was therefore omitted for that speaker. Due to these unforeseen circumstances, the data were not completely balanced for all controls (manner, place, voice type, vowel quality) across all speakers. Notably, there was an imbalance found in the aspirated palatal stops (about 8 tokens for all speakers except speaker KT), and the aspirated retroflex stops (about 12 tokens for all speakers). The total number of tokens for each control is given per speaker is given in Table 6.

In total, 730 tokens were analysed for KT, 720 for CH, 696 for KL, 707 for DM, and 687 for SD.^[14] The minor discrepancies in total tokens recorded per speaker should not have a tangible effect on the results of this study. The larger imbalances found in the retroflexes and palatals are discussed in relation to their results.

Recordings done in 2009 were made using a Shure brand, head mounted microphone, placed approximately 3 cm from the speaker’s mouth using a Marantz PMD 660 recorder. A Zoom H4N recorder was used for the recordings of the other three speakers (KL, SD, and DM), using the internal microphone. We opted against an external microphone with this recorder as it was found to actually reduce the quality of the recording.^[15] Both recordings were made at a 24-bit 96 kHz sampling rate as .wav format. Acoustic analysis was completed using Praat (Boersma and Weenick 2014) acoustic software.

2.3 Acoustic measurements

VOT was measured by hand between the release and the first voicing cycle (Lisker and Abramson 1964). Occasionally frication, particularly at the palatal and velar places of articulation, was included in the VOT measurements of the stops. The effects of this are considered in the discussion. Voicing for fricatives was determined primarily by the presence of a voice bar in the spectrogram, glottal pulsations in the spectrogram and often in the wave form, and periodic wave forms (as opposed to the aperiodic voiceless wave form). Duration was measured for the fricatives since Stevens et al. (1992) showed that duration must exceed 60 ms for voicing (particularly relating to voicelessness) to be perceived.^[16] Vowel duration was measured by hand according to guidelines in Wright and Nichols (2009).

The fundamental frequency (f0) was measured at nine equidistant points on the vowel using a Praat script.^[17] Since the voicing and sonority of the prevocalic consonant can have a contour effect on the pitch levels (Hombert et al. 1979), the approximate mid-point (interval 4) was used in statistical analysis. Fundamental frequency at the onset (interval 1) and the end of the vowel (interval 9) were also statistically assessed to test if the contrast was maintained across the duration of the vowel.

3.1 Fundamental frequency

Figures 2 –6 represent the mean f0 of the vowel following the voiced and voiceless dental fricatives for each of the five speakers recorded.

Figure 2:

Mean f0 from 75 tokens of vowels following dental fricatives for speaker CH.

Figure 3:

Mean f0 from 72 tokens of vowels following dental fricatives for speaker DM.

Figure 4:

Mean f0 from 72 tokens of vowels following dental fricatives for speaker KL.

Figure 5:

Mean f0 from 72 tokens of vowels following dental fricatives for speaker KT.

Figure 6:

Mean f0 from 72 tokens of vowels following dental fricatives for speaker SD.

Figures 2 –6 show that the difference in mean f0 following the voiced and voiceless series is maintained across the duration of the vowel for all speakers, with the exception of speaker SD. Figure 6 shows that for SD the difference between mean f0 appears to taper towards the end of the vowel (from around 70 %). To query whether the mean f0 following the voiceless dental fricative and that of the voiced dental fricative are distinct categories representing tone (high and low respectively) a linear mixed model was run, with f0 at interval 4 as the outcome, speaker as the random effect, and voice type (voiced or voiceless) as the fixed type. Figure 7 shows a boxplot of the mean f0 at interval four for voiced versus voiceless fricatives, by speaker. Note also that the variation in pitch appears much smaller following the voiced than the voiceless category; this would be expected if the primary contrast of the voiced category had moved into the domain of tone.

Figure 7:

Clustered boxplot of mean f0 at interval four by speaker and voice type.

The estimated means are shown in Table 7; note that the f0 following phonologically voiceless fricatives is nearly 30 Hz higher than that of the phonologically voiced fricatives.

Tables 8 and 9 show that this difference in interval 4 f0 is statistically distinct for the two categories.

3.2 Voicing

We measured closure duration of each fricative and also coded each for presence or absence of voicing. Stevens et al. (1992) reported that voicing in fricatives can only be perceived if the duration is more than 60 ms. In our study, Kurtöp voiceless fricatives had an average closure duration of 143 ms while the closure duration of voiced fricatives was 131 ms; thus there is ample duration to encode a voicing contrast. Voicing was determined by the presence of a voice bar in the spectrogram, glottal pulsation in the spectrogram and wave form, and aperiodic versus periodic wave forms. This was usually an easy and straightforward process, illustrated by the difference shown in Figures 8 and 9 below.

Figure 8 (L) and 9 (R):

Voiced and voiceless realisations of Ch’s pronunciation of zi ‘cat eye’. Note the salient voice bar associated with the fricative in 8, but completely absent in 9. 8 also shows a periodic wave form going into the vowel while 9 shows only aperiodic noise prior to the onset of the vowel.

For the vast majority of tokens, the decision to label a fricative as voiced or voiceless was straightforward. Very occasionally, the voicing would be less easy to determine, as in Figure 10. Here, the beginning of the wave form appears periodic but then is aperiodic for the majority of the closure duration. This voicing could be due to the fact that this token was the iteration from the carrier phrase; thus the voicing at the beginning could be the trailing off of the preceding vowel (as in Figure 10). Even more rarely, voicing sometimes started halfway through the fricative and before the start of the vowel, as in Figure 11. In these ambiguous cases, we labelled a sound as voiced if more than 50 % of the duration was voiced while it was labelled voiceless if less than 50 % of the duration was voiceless. In reality, these ambiguous cases comprise a very small proportion of the data set; for example, of KT’s token’s, only two were somewhat ambiguous; these are shown in Figures 10 and 11.

Figure 10 (L) and 11 (R):

KT voiceless iteration of zon ‘two’ and voiced iteration of ze ‘substance’.

By isolating the phonemically voiced dental fricatives, we can look at percentage of tokens with the presence of voicing (voiced) versus without voicing (voiceless). Overall, the phonemic voiced dental fricatives have a larger portion (62 %) of voiceless tokens (lacking the presence of a voice bar or glottal pulsations) than voiced tokens (38 %). This suggests that the voice distinction in the voiced dental fricatives is merging with the voiceless dental fricatives at a faster rate than for the stops.

Looking at the individual speakers in terms of percentages of total tokens with the presence of a voicing (voiced) and without voicing (voiceless) as in Figure 12, it becomes clear that two speakers (DM and KT) have nearly lost the voicing distinction in the dental fricatives.

Figure 12:

Percentage of total phonetically voiceless and voiced tokens for each speaker for phonemically voiced dental fricatives.

Figure 12 shows that for DM around 95 % of voiced dental fricatives are realised as voiceless. For speaker KT this is closer to 80 %. Over 60 % of CH’s tokens are realised as voiceless, while speaker KL appears to be just the opposite. The only other speaker, besides KL, to have more tokens realised as voiced than tokens realized as voiceless is speaker SD with a nearly 50/50 distribution of his phonemically voiced fricatives.

4.1 Stops as a whole

Prior to getting into the detailed results for each place of articulation, we will look briefly at the stops as a whole. The results in general corroborate the results found in Hyslop (2009), with minor differences. The low tone following voiced stops and the high tone following voiceless stops are maintained at a statistically significant difference across the entire duration of the vowel and the voiced categories of stops is merging with the voiceless category.

4.1.1 F0 on following vowels

We see that the f0 following the voiceless stop and the f0 following the aspirated stop are significantly distinct for all speakers at the near midpoint of the vowel (interval 4). When looking at the onset and end positions of the vowel, the mean f0 following the voiceless and aspirated stops is not statistically significant for one speaker (KT) at the onset (interval 1), and for three speakers (CH, KT SD) at the end of the vowel (interval 9). The means and standard deviation of the VOT for all three voice types (voiced, voiceless, aspirated) were proven to be statistically significant categories for all five speakers. The standard deviation of the voiced series was much higher than the other two voicing categories, which supports Hyslop’s (2009) proposal that there is an ongoing merger between the voiced and voiceless series and thus there is a great deal of variation in realisation of VOT within the phonemically voiced category.

Figures 13–17 show the f0 on the vowel following voiced, voiceless, and voiceless aspirated stops for each speaker.

Figure 13:

Mean f0 from 645 tokens of vowels following stops for speaker CH.

Figure 14:

Mean f0 from 635 tokens of vowels following stops for speaker DM.

Figure 15:

Mean f0 from 624 tokens of vowels following stops for speaker KL.

Figure 16:

Mean f0 from 658 tokens of vowels following stops for speaker KT.

Figure 17:

Mean f0 from 615 tokens of vowels following stops for speaker SD.

We see a difference in f0 on vowels with a lower f0 following the voiced stops and a higher f0 following the two voiceless series (voiceless and voiceless aspirated). For speakers KL and DM there appears to be a difference in high f0 between the aspirated and voiceless series; however, this difference is less obvious for the remaining three speakers (CH, KT, and SD).

At a glance, the f0 differences appear substantial for all speakers at all three points in the vowel. Note that for all three points (onset, mid point, end point), KT exhibits the lowest amount of difference between means. Interestingly, as noted visually from Figure 13, CH has the highest degree of difference at the onset of the vowel and the third lowest at the end of the vowel, which suggests that CH exhibits the highest degree of variation in mean f0 differences across the duration of the vowel. All of these differences will be explored through the following statistical analysis.

The mean f0 measured at the midpoint of the vowel (interval 4) increased following voiced, aspirated, and voiceless stops, in that order, for all five speakers. The means are given in Table 10, showing that mean f0 following voiced stops were consistently the lowest across all the speakers.

Table 10:

Summary of mean F ₀ following stops at midpoint of the vowel.

Speaker	Voice type	N	Mean	S.D
CH	Voiced	236	258.84	15.35
	Aspirated	157	295.66	17.51
	Voiceless	252	306.88	21.99
DM	Voiced	236	217.72	9.05
	Aspirated	163	251.25	27.89
	Voiceless	236	268.29	22.77
KL	Voiced	228	132.91	6.58
	Aspirated	160	162.98	13.87
	Voiceless	240	175.90	17.39
KT	Voiced	247	117.84	8.88
	Aspirated	187	140.77	14.09
	Voiceless	224	145.75	14.36
SD	Voiced	232	158.46	11.84
	Aspirated	152	188.69	13.35
	Voiceless	231	195.81	16.57

1. Note. Number of tokens, mean and standard deviation for F ₀ on the midpoint (interval 4) for all speakers within each voicing type.

Statistical analysis confirmed the observation above via a Linear Mixed Model; results are shown below in Table 11, examining f0 (interval four). The model is fitted for an interaction of term between voice type and place of articulation, keeping subject as a random effect.

Table 11:

Linear mixed model results.

Fixed effectsa
Source	F	dF1	Df2	Sig.
Corrected model	353.096	14	3,165	0.000
VoiceType * PoA	17.314	8	3,165	0.000
VoiceType	1991.088	2	3,165	0.000
PoA	28.002	4	3,165	0.000

1. Probability distribution: Normal. Link function: Identity^a. ^aTarget: Pitch 4.

We can confirm that the mean f0 following stops exhibits statistically significant categories that are connected to the voice type of the preceding stop. For further tests at different places across the vowel, the reader is referred to Plane (2016).

Figure 18 is a visual representation of the mean f0 in interval four as a function of voice type; note again that we see more variation in f0 following voiceless and aspirated stops than following the voiced category. We will return to this issue later.

Figure 18:

Mean f0 in interval four as a feature of phonological voicing type.

4.1.2 VOT

Mean and standard deviation of VOT for voiced, voiceless, and voiceless aspirated tokens were also calculated for all speakers and the results are given in Table 9.

Looking at Table 12, mean scores of all three categories appear quite different for all speakers. For every speaker, the mean increases according to voice type, with voiced being the lowest and aspirated being the highest. Notice that DM’s mean score for voiced series (−103.60) is well below any of the other speakers, yet DM’s voiced series range (−257 ms to 83 ms) is similar to CH’s and SD’s voiced series range [CH (−250 ms to 96 ms) and SD (−250 ms to 110 ms)].

Table 12:

VOT summary for stops for all speakers.

Speaker	Voice type	N	Mean	S.D	Range
CH	Voiced	236	−48.01	58.15	−250 to 96
	Voiceless	252	16.15	10.89	1–65
	Aspirated	157	40.96	18.05	6–91
DM	Voiced	236	−103.60	84.0	−257 to 83
	Voiceless	236	33.0	25.21	3–112
	Aspirated	163	68.11	31.51	4–156
KL	Voiced	228	−17.58	54.83	−154 to 55
	Voiceless	240	21.55	14.06	6–77
	Aspirated	160	57.26	18.87	24–102
KT	Voiced	247	−21.64	62.86	−168 to 84
	Voiceless	224	31.0	15.06	3–80
	Aspirated	187	80.0	23.76	30–154
SD	Voiced	232	−40.83	56.42	−250 to110
	Voiceless	231	21.71	16.38	1–68
	Aspirated	152	55.85	21.85	18–150
Total	Voiced	1,179	−46.3	71.14	−257 to 110
	Voiceless	1,190	24.51	18.09	1–112
	Aspirated	819	61.26	26.88	4–156

1. Note. Number mean standard deviation and range values are shown for each speaker and each voice type.

Statistical analysis confirmed that all three voice types are statistically significant categories for syllable-initial stops. Homogeneity of variance was violated, as assessed by Levene’s Test of Homogeneity of Variance (p < 0.001) for all speakers. Welch’s ANOVA confirmed the statistical significance for all speakers: for speaker CH [F(2,313.235) = 280.884, p < 0.001], for speaker DM [F(2,361.272) = 411.688, p < 0.001], for speaker KL [F(2,338.836) = 294.567, p < 0.001], for speaker KT [F(2,373.547) = 424.548, p < 0.001], and for speaker SD [F(2,335.578) = 311.489, p < 0.001]. A Games-Howell post hoc test confirmed (p < 0.001) the statistical significance for all speakers for each of the three paired comparisons of the three voicing types.

Turning now only to the phonologically voiced category, we can see the individual speaker differences in Figure 19.

Figure 19:

Phonologically voiced stops with positive and negative VOTs for each speaker.

Figure 19 shows that speaker CH has a 60 % negative, 40 % positive distribution of stops, identical to the overall results. Both DM and SD have a larger percentage of negative tokens (80 % and 70 % respectively). Speaker KT is nearly split 50/50, while speaker KL has the reverse of CH and the overall results shows 40 % negative and 60 % positive tokens.

The scatter plot in Figure 20 represents visually distinct categories between the low f0 following voiced stops and the higher f0 following the voiceless and aspirated stops, with their correlation to VOT.

Figure 20:

Relationship between VOT and f0 following stops for each speaker.

The blue series represents the voiced stops. Note that they occupy the left side of the graph (the negative VOT), while still blending with the voiceless stops (green series) on the positive VOT portion of the graph. Also note that while the series merges on the positive side of the VOT, the voiced series is lower in f0 than the voiceless series. The voiceless series appears to have a larger f0 range, while the voiced stops seem to have a narrow f0 range with a larger VOT range.

4.2 Place of articulation differences

With the merger between voiceless and voiced stops in favour of tone clearly underway, we can now turn to the differences for each place of articulation. To determine if there could be a trading relationship between the devoicing of a phonologically voiced stop and pitch on the following vowel, we plotted f0 in interval four, (on the y-axis) as a function of VOT (on the x-axis) for each voicing category for each speaker in Figures 21–25.

Figures 21–25:

f0 and VOT for the stops varying in voicing category and place of articulation for each speaker.

There are several things to note in the above figures. First, if we look at the distribution of the voiceless (green) and aspirated (pink) VOT values, we see the expected distribution on the right side of the X-axis, showing positive VOTs. Further, despite some overlap, the two categories have a fairly uniform f0 distribution. We see more speaker variation within the phonologically voiced series. Blue dots to the left of the 0 indicate pre-voicing and those blue dots produced to the right of the 0 are produced as voiceless acoustically.

We can make a further observation regarding the phonologically voiced category. When these tokens are realised with pre-voicing, we see a large degree of variation in the realization of f0 on the midpoint of the following vowel. For example, f0 on the following vowels of pre-voiced initials for Ch (upper left) ranges from below 250 Hz to approximately 350 Hz and f0 on the following vowels of pre-voiced initials for speaker KT (middle left) ranges from just below 100 Hz to approximately 175 Hz. On the other hand, when these phonologically voiced tokens are realised as voiceless, the f0 variation on the following vowel disappears. This is true for all speakers, but perhaps most pronounced for KL (middle right). If we look in particular at KL’s retroflex tokens that are phonetically voiceless, we see that the f0 space for each different voicing type (voiceless, aspirated, voiced) is distinct. KL’s phonetically voiceless but phonologically voiced retroflexes have a coherent f0 on their following vowel with just below 150 Hz, whereas the f0 on the vowels following the phonologically voiceless retroflexes is closer to 200 Hz. This is true for all speakers and the very strong evidence of f0 correlations discussed above can be taken to be directly related to the devoicing of the phonologically voiced series. Note, importantly, that we sometimes get low f0 without concomitant devoicing. That is, we can have phonetically voiced segments with low ensuing pitch, but we always have low pitch following a phonetically devoiced segment. This suggests that pitch phonologises before the voicing is lost.

Figures 21–25 also illustrate the drastic differences between the retroflexes and labials. If we take KL’s (middle right) retroflexes, for example, we see that the voiceless and aspirated categories have fairly uniform VOT realisations together with distinct f0 realisations on the following vowel. There are only a few phonetically voiced retroflex tokens; all of these have low f0 realised on their following vowels.

On the other hand, we can look at labials for speakers such as CH, KT, and SD, and notice the preponderance of acoustically voiced realisations and with f0 realisations on following vowels that appear to span the extent of the f0 space for each speaker. There are some voiceless realisations, all with low f0 on the following vowel, but these are a minority of the tokens. The dentals are similarly more frequently voiced and have a range of f0 realisations on following vowel. We can also see in these figures that the palatals and velars appear to be an intermediate category between the retroflexes and labials/dentals. Interestingly, we also see what appears to be a near bimodal distribution in the f0 realisation following the phonemically voiced category of velars for KT and, slightly so, for DM.

Given the expectation that phonologically voiced stops would be devoicing in favour of a tonal contrast, we now turn to the voiced results alone. Figure 25 provides the percentages of tokens that have positive VOT and those that have negative VOT in the voiced series, organized by each speaker at each place of articulation. Figure 26 also shows several noteworthy developments. First, CH and DM do not have any places of articulation with a higher percentage of positive VOT than negative VOT. Second, KL has more positive than negative VOT tokens in all places of articulation except labial. Third, going from front to back, the number of speakers that have more positive VOT than negative VOT are as follows: labial has zero of five speakers, dental has one of five speakers (with KT approaching 50/50), retroflex has three of five speakers (with speaker CH approaching 50/50), palatal has two of five speakers (with SD approaching 50/50), and velar has two of five speakers.

Figure 26:

Percentage of total positive and negative VOT tokens for each speaker at each place of articulation.

4.2.1 Labial

The histogram in Figure 27 shows the VOT of the voiced labial stops for each speaker.

Figure 27:

Distribution of VOT for voiced labial stops for each speaker.

There is a bimodal distribution in stops for most of the speakers (speakers DM and SD both have some positively realised labial stops but have less pronounced modes compared to the other speakers). The above data show that the voiced labial stops have both negative and positive VOT for all speakers. In order to test whether or not there is a trade off between phonetic voiced and pitch on the following vowel, we ran a linear mixed model, keeping speaker as a random effect. The estimates are shown in Table 13. These two categories are not statistically significant (p = 0.336).

Table 13:

Estimates for interval 4 f0 as a function of phonetic voicing.

Estimates
VOT binary	Mean	Std. Error	95 % confidence interval
			Lower	Upper
0	175.712	43.408	90.377	261.047
1	174.499	43.417	89.146	259.851

4.2.2 Dental

Figure 28 provides visual representation of the VOT distribution of the voiced dentals.

Figure 28:

Distribution of VOT for voiced dental stops for each speaker.

The means for voiced dental stops are negative for all speakers; however, a few speakers have smaller negative means than those found in the labial stops. Both KT’s voiced dental mean (−37.75) and KL’s (−8.08) are smaller than the smallest voiced mean in the labial series (where the smallest was speaker KT at −44.51). The range column shows that there is overlap between the voiced and voiceless series (as well as the voiceless and aspirated).

It appears that fewer speakers have a clear bimodal distribution than seen in the labials. DM and SD both have few positive tokens. KT appears to have a similar bimodal distribution as in the labials, but perhaps slightly more skewed towards the positive end in the dentals. CH has a less pronounced mode in the positive end than was represented by the labials. One speaker (KL) has a much more pronounced mode in the positives than in the negatives.

In order to ascertain whether or not there is a trade off between phonetic voicing and pitch on the following vowel, we ran a linear mixed model, keeping speaker as a random effect. The estimates are shown in Table 14. These two categories are not statistically significant (p = 0.206).

Table 14:

Estimates for interval 4 f0 as a function of phonetic voicing; 0 = phonetically voiced; 1 = phonetically voiceless.

Estimates
VOT binary	Mean	Std. Error	95 % confidence interval
			Lower	Upper
0	178.945	41.972	96.153	261.736
1	176.542	41.988	93.720	259.364

4.2.3 Retroflex

Isolating the voiced series and looking at the distribution in Figure 29, it appears that for speakers KL, KT, and SD, the voiced retroflexes have almost completely merged with the voiceless series. This is not surprising, as it was already discussed that there was not a statistical difference between these two series for two of these speakers (KL and SD).

Figure 29:

Distribution of VOT for voiced retroflex stops for each speaker.

In order to ascertain whether or not there is a trade off between phonetic voicing and pitch on the following vowel, we ran a linear mixed model, keeping speaker as a random effect. The estimates are shown in Table 15. This time the difference is statistically significant (Table 16), suggesting that tone on the following vowel, rather than voice, is the primary contrast amongst retroflex stops.

Table 15:

Estimates for interval 4 f0 as a function of phonetic voicing; 0 = phonetically voiced; 1 = phonetically voiceless.

Estimates
VOT binary	Mean	Std. Error	95 % confidence interval
			Lower	Upper
0	181.727	38.946	104.921	258.532
1	176.395	38.937	99.607	253.182

Table 16:

Statistical significance of f0 at interval four, as a function of phonetic voicing.

Fixed effectsa
Source	F	dF1	Df2	Sig.
Corrected model	12.450	1	197	<0.001
VOTbinary	12.450	1	197	<0.001

1. Probability distribution: Normal. Link function: Identity^a. ^aTarget: Pitch 4.

4.2.4 Palatal

Figure 30 offers an isolated look at the voiced palatal series for all speakers.

Figure 30:

Distribution of VOT for voiced palatal stops for each speaker.

Again there is a positive mean in the voiced series. For speakers DM and SD, the voiceless and aspirated series were not distinct categories, though this could be a result of the small sample size of the aspirated series (8 tokens for all speakers except KT who had 40 tokens). Focussing on the voiced series, KL has the smallest standard deviation (52.82) and DM has the largest (108.75), among any of the places of articulation, thus far (108.75). DM also has the largest negative mean of the voiceless series.

Looking at Figure 30, all speakers have some tokens that are voiced with positive VOT. KT and SD have distinct bimodal distributions. CH and DM have less pronounced modes in the positive range, but still appear to have a bimodal distribution. KL appears to favour the positive VOT realisation and the few tokens that are realised with negative VOT appear to be the exception. This suggests that KL has nearly merged the voicing distinction between voiced and voiceless palatal stops.

In order to ascertain whether or not there is a trade-off between phonetic voicing and f0 on the following vowel, we ran a linear mixed model, keeping speaker as a random effect. The estimates are shown in Table 17. The difference is statistically significant, as shown in Table 18.

Table 17:

Estimates for interval 4 f0 as a function of phonetic voicing; 0 = phonetically voiced; 1 = phonetically voiceless.

Estimates
VOT binary	Mean	Std. Error	95 % confidence interval
			Lower	Upper
0	181.783	38.913	104.840	258.726
1	175.730	38.917	98.779	252.681

Table 18:

Statistical significance of f0 at interval four, as a function of phonetic voicing.

Fixed coefficientsa
Model term	Coefficient	Std. Error	t	Sig.	95 % confidence interval
					Lower	Upper
Intercept	175.730	38.9170	4.516	<0.001	98.779	252.681
VOTbinary = 0	6.053	2.2788	2.656	0.009	1.548	10.559
VOTbinary = 1	0b	-	-	-	-	-

1. Probability distribution: Normal. Link function: Identity. ^aTarget: Pitch 4. ^bThis coefficient is set to zero because it is redundant.

4.2.5 Velar

Figure 31 presents the distribution of the voiced series VOT.

Figure 31:

Distribution of VOT for voiced velar stops for each speaker.

The distribution of the voiced velar series in Figure 31 shows a considerable trend towards positive VOT realization of these stops. KT and KL once again have merged more of their tokens with the voiceless series, while SD, DM and CH appear to have more voiced VOT token but also have sizable representation of the merger with the voiceless series.

In order to ascertain whether or not there is a trade-off between phonetic voiced and pitch on the following vowel, we ran a linear mixed model, keeping speaker as a random effect. The estimates are shown in Table 19. The difference is not statistically significant (p = 0.198).

Table 19:

Estimates for interval 4 f0 as a function of phonetic voicing; 0 = phonetically voiced; 1 = phonetically voiceless.

Estimates
VOT binary	Mean	Std. Error	95 % confidence interval
			Lower	Upper
0	176.257	38.849	99.731	252.783
1	178.633	38.853	102.099	255.167

5.1 Voiced obstruents

Given that we argue the voiced obstruents are currently undergoing tonogenesis and merging with the voiceless series, in favour of a tonal contrast following vowels, it is worth considering the voiced results in detail; the issue of cue-trading in phonologisation becomes especially relevant at this stage in tonogenesis. We have shown that both voicing and f0 are utilised in making a contrast amongst Kurtöp obstruents and within the phonologically voiced category there is a particularly high variation in how the ‘voicing’ is realised. We further argue that the devoicing of the obstruents is motivated by the transphonologisation of f0 as tone on the following vowel.

In order to more confidently assert this, we have plotted f0 at the approximate mid point (interval 4) as a function of presence or absence of voicing amongst the phonologically voiced stops. Due to the high amount of variation and relatively small numbers of tokens for some categories for some speakers, we examined f0 (interval four) as a function of phonetic voicing for each speaker individually, rather than aggregating across all speakers. These results are shown in Figure 32. Unfortunately, the small number of tokens and high variability amongst DM in particular means a comparison of fricatives versus stops as a whole is not possible. Nonetheless, we can see for at least speakers SD and KL, there is less variation in pitch with the phonetic voiceless realisations.

Figure 32:

f0 as a function of voicing across phonologically voiced fricatives for each speaker. Error bars represent confidence interval for mean; 0 = phoneticlly voiced; 1 = phonetically voiceless.

We also plotted f0 at interval four as a function of phonetic voicing amongst stops, looking at each place of articulation separately. These results are summarised for all speakers in Figure 33.

Figure 33:

f0 as a function of voicing across phonologically voiced stops for each place of articulation across all speakers; 0 = phonetically voiced; 1 = phonetically voiceless.

Examining approximate mid point f0 on vowels following phonetically voiced versus phonetically voiceless stops for each place of articulation confirms the observation made above; tonogenesis appears to be happening for some places of articulation at a faster rate than for others. Specifically, we see that f0 is lower following phonetically voiceless retroflexes, palatals, velars and dentals, while pitch following labials appears relatively equal, regardless of whether or not there is phonetic voicing.

These figures further illustrate the cue-trading relationship between VOT and f0 as part of the tonogenesis process. Schertz and Clare (2020: 5) argue that when change is in progress, some speakers will rely more on the ‘innovative’ cue while others rely more on the traditional cue. In our case, the ‘traditional’ cue is VOT and the innovative one is f0. We can see that speakers do indeed do different things with these cues, but also, crucially important here, that the cues are utilised differently at different places of articulation.

5.2 Obstruents as a whole

These results have shown us several things. First, f0 is statistically higher following phonologically voiceless obstruents than when following phonologically voiced obstruents, fitting in with the received model of tonogenesis in which voiceless onsets condition high tone on following vowels and voiced onsets condition low tone on following vowels. Second, we have seen that phonologically voiced obstruents are very often realised as phonetically voiceless, again following the expected tonogenetic developments.

Considering these results from a lens of transphonologisation, we see that while both f0 and voicing (either as VOT or presence of voicing during fricative closure) are both still present as part of the acoustic bundle of features marking a phonemic contrast in Kurtöp, there appears to be a shift in importance. F0 is statistically significant on the following vowels overall, suggesting that this is a primary feature of the contrast in production. However, the voicing results showed high variability, suggesting that for some place/manners of articulation voicing is a more prominent feature of the acoustic bundle than for other places. In places where voicing is less important, we make the inference that tone has become primary and thus devoicing can now happen. The transphonologisation from a contrast in voice on obstruents to contrast in tone on following vowels is still underway for all obstruents but palatal fricatives. However, the different weights attributed to the different cues for the different places and manners strongly suggests transphonologisation will be completed for some places and manners before others.

Stepping back to look at the bigger picture, we can fit the fricatives into the larger tonal pathway of the language. As has been established, tone has first entered Kurtöp following the complex onsets in which the second member was a nasal consonant. With tone contrastive following the nasals, predictable pitch must have developed following obstruents, in particular the voiceless and voiced palatal fricatives.

The results presented in Section 3 and 4.1 suggest that the fricatives are further along in the process of transphonologisation than the stops as a whole. Specifically, Section 3 showed that there is a statistically significant distinction in mean f0 following the voiced dental fricatives (low) and the mean f0 following the voiceless dental fricatives (high). These results suggest that high and low tones on the vowel are distinct categories following the voiceless and phonologically voiced dental fricatives, respectively. The results were statistically significant for three speakers (DM, KL, KT) in all three tested points along the vowel (onset, midpoint, and endpoint), and were statistically significant for two of the speakers (SD and CH) at the onset and midpoint, but not at the endpoint. In particular, the visual representation of SD’s mean f0 was atypical compared to the other speakers. It is unclear what motivations caused SD’s results to be so incongruous with the other speakers.

The evidence presented here supports the prediction by Hyslop (2009) that all obstruents (not just stops) are merging the voicing distinction in favour of a tonal contrast. Further, the fricatives are merging more quickly than the stops. This is summarised in Figures 34 and 35, showing that phonologically voiced stops are devoiced approximately 40 percent of the time while phonologically voiced fricatives are devoiced over 60 percent of the time.

Figure 34 (L) and 35 (R):

Percentage of phonetically voiceless versus voiced realisations of phonologically voiced stops (left) as opposed to fricatives (right).

Kurtöp’s five places of articulation provide an ideal location to query the role of place of articulation in the propagation of tone through a language. Overall, we saw that speakers favoured devoicing of the retroflexes over the other four places of articulation. We are unsure precisely why this is the case; however, we can put forward some hypotheses. As Hyslop (2014, 2017 outlines, Kurtöp retroflexes have recently entered the language through the simplification of velar plus rhotic onset clusters.^[18] In fact, for some speakers they might be better described as doubly-articulated segments, often with a strong, separate rhotic realisation. A summary of words showing the Tibetan source of retroflexes in Kurtöp is shown in Table 20.^[19]

Given that tone first entered in Kurtöp following complex onsets and that complex onset simplification also seems to be happening as a slow diachronic process in Kurtöp Hyslop (2017), it seems possible that the retroflexes are still phonologically – in some senses – complex onsets (note also that they do not occur in coda position). Hence, if we re-class retroflexes as complex onsets with sonorant second members, we might expect them to behave more akin to the complex onsets with nasal second members and thus be one of the first places to condition the phonologisation of tone. Additional hypotheses to explain the preference to devoice retroflexes include intrinsic closure duration differences^[20] and social factors.^[21]

Following the retroflexes, we found that the velars and then the palatals were next likely to devoice. There is, perhaps, a physiological explanation for this. Ohala and Riordan (1979) and Ohala (2012) point out that it is harder to maintain voicing in velars than it is in bilabials, a fact which is mirrored in phenomena such as the ‘velar gap’, a tendency for a voiced consonant, when missing from an otherwise balanced stop inventory, to be a voiced velar. Conversely then, we find that dentals and especially bilabials are much less likely to devoice than the other places of articulation. More research is needed to explain why back places of articulation would be more likely to devoice than front places, perhaps through the physiological motivations mentioned above.

5.3 Typological considerations

The Kurtöp findings can be compared and contrasted with other languages. Tone is reported to enter Tai (Pittayaporn 2009) and Tibetan (Mazaudon 1977) following the most sonorous segments first. Rhotics alone have conditioned tonogenesis in the Phnom Pehn dialect of Khmer tonogenesis (Wayland and Guion 2005). Hyslop (2022) presents a typology of tonogenesis, showing that in fact such importance on sonorants in the development of tone is typologically more common than previously expected.

Afrikaans (Coetzee et al. 2018) also has incipient tonogenesis. Due to the collapse of a VOT contrast in initial position, f0 has increasingly been observed to carry more weight perceptually and is also now being produced by speakers. Note that Coetzee et al. (2018) also mention that the labial place of articulation, in particular, relies more on voicing in perception for older speakers. Shryock’s (1995) acoustic study of the Chadic language Musey finds that overall, stops are more often produced as voiceless than fricatives, and non-laryngeal fricatives are more often voiced than [h] (17). Fricatives are also reported to play a privileged role in tonogenesis described for Athabaskan (Kingston 2007), Balsas Nahuatl (Guion et al. 2009) and Kickapoo (Gathercole 1983). Thus there is some typological evidence to support the notion that fricatives may be more ‘prone’ to tonogenesis than stops, and that amongst the stops, labials may be the last to devoice.