Flexibility and evolution of cue weighting after a tonal split: an experimental field study on Tamang

1.1 Tamang and the linguistic situation in Nepal

Nepal is home to around 130 languages of Indo-Aryan, Tibeto-Burman, and other families, some tonal, most toneless. With over 1,300,000 mother-tongue speakers (Nepal census 2011), the Tamang community constitutes the largest non-Indo-Aryan group in Nepal. Until the mid-twentieth century, the Tamang, mostly endogamous and village-dwellers, were largely monolingual. A few Tamang were literate in Nepali, the Indo-Aryan national language, or in Tibetan. Very occasionally, Tamang was jotted down for private use in one of the corresponding alphabets, Devanagari or Tibetan. However, in the last 40 years, bilingualism and schooling in Nepali have become widespread.

Tamang belongs to the Tamangish group (also known as TGTM or Tamang-Gurung-Thakali-Manangke), a branch of the Bodish section of Tibeto-Burman which also contains Tibetan, as well as Tshangla (ISO [tsj]) and Kurtöp (ISO [xkz]) in Bhutan. There are several dialects of Tamang. The current study bears on Risiangku Tamang, an Eastern Tamang variety (ISO [taj]). The administrative unit of Risiangku (alternative names: Lisankhu, Risingo; 27°6′ N, 85°8′ E) has 3,700 residents, 2,300 of whom are Tamang (Nepal census 2011). Risiangku is recognized by the Tamang as a cultural center, boasting a Buddhist temple since AD 1093 (Lama 1959).

1.2 The Tamang tonal split in an areal context

All modern TGTM languages are tonal, with systems of four tones defined on lexemes, which are largely monosyllabic (almost all verb roots and half the nouns). Suffixes have no independent tones and the tonal characteristics of the initial morpheme extend over the whole phonological word, constituting a word-tone system, first described in Pike (1970) and Mazaudon (1973). Proto-TGTM, the ancestor of all TGTM languages, has been reconstructed with three plosive series (voiceless aspirated, voiceless unaspirated, and prevoiced), two series of continuants (voiceless and voiced fricatives and sonorants), and only two tones (Mazaudon 1978, 2012). The gradual loss of the voicing distinction on initial consonants was the source of the split of the old two-tone system into the modern four-tone systems. While some residual voicing of initial plosives survives in some of the languages, the two series of continuants have completely merged in all TGTM languages, thus finalizing the phonemicization of the new tonal contrast.^[1]

The consonantal mutation and tonal split which we presently observe in Tamang is an instantiation of the vast Asian mutation which started more than a millennium ago in Middle Chinese and swept through most parts of Asia in various forms (Maspero 1912; Haudricourt 1961, 1965). Schematically, the loss of a voicing contrast on initial consonants was replaced by two orders of contrasts: (1) in previously tonal languages, like Middle Chinese, Vietnamese, or Thai, the number of tones was doubled or sometimes tripled; (2) in previously toneless languages, like Khmer, a register contrast was created, usually manifested by a doubling of vowel timbres, often associated with secondary features of breathiness, pitch, or vowel length (Ferlus 1979; Huffman 1976).

Multiple cues are known to be involved in register languages, and they have been studied experimentally (for a review see Brunelle and Kirby 2016). But tonal languages are not pure pitch either. Haudricourt (1965) proposed that at the beginning of a tonal split, breathy voice often accompanied the devoicing of the old voiced plosives. Breathiness was retained alongside some voicing of initial consonants in the Wu dialects of Chinese (for recent experimental studies, see, e.g., Gao 2015; Tian and Kuang 2021; Jiang et al. 2020; Zhang and Yan 2018).

This stage can be observed in the Tamangish languages, where multiple cues contribute to the tonal contrasts but their realizations and relative weights differ between varieties. In conservative varieties, there is a clear high versus low tonal contrast: the two high tones, produced with high pitch onset and modal voice, co-occur with aspirated and unaspirated voiceless plosives, while the two low tones, produced with low F0 and breathy voice, co-occur with a single series of plosives, unaspirated, occasionally realized with partial or full voicing. In evolved varieties, such as Manangke, Marphali, or Taglung Tamang, further changes have obscured the etymological picture (Hildebrandt 2005; Gao and Mazaudon 2017; Mazaudon 1978).

1.3 Summary of production data in Risiangku Tamang

To better understand the synchronic relationships between features that arose from the tonal split, an earlier experimental investigation was conducted on one of the conservative varieties, Risiangku Tamang, using electroglottographic data from five male speakers in their 30s–40s (Mazaudon and Michaud 2008; Michaud and Mazaudon 2006). The phonetic description of the four tones is summarized as follows:

1. F0 height and contour: T1 is the highest and has a falling contour; T2 is the second highest; T3 is low and rising; T4 is the lowest and falling.

2. Phonation: T1 and T2 are modal; T3 and T4 are breathy (or “whispery” as used in that study) but T4 to a lesser degree. On average, glottal open quotient is above 50% for T1 and T2 for all speakers, and higher (from 55 to 71%) for T3 and T4, which suggests that the phonation difference is between modal and breathy, rather than, say, laryngealized/tense and lax (Maddieson and Ladefoged 1985).

3. Voicing: 20–30% of the initial plosives of low tone words are fully or partially voiced, as opposed to almost 0% in the high tone words.^[2]

In the Tamang word-tone system, the F0 contour extends over the entire word but noninitial syllables are subject to intonational variability, as shown in Figure 1, which plots the mean F0 curves in semitones of a limited number of words (n = 43) produced by the male participants of the current study. Figure 2 shows the mean F0 curves in hertz of the first syllable alone. The F0 curves are similar to the plots in Michaud and Mazaudon (2006) and Mazaudon and Michaud (2008), which can be consulted for more details, as well as Mazaudon (1973) and Mazaudon (2005: 84, Figure 4).

Figure 1:

Mean smoothed F0 curves in semitones of the four tones produced by the male participants of the current study: disyllabic /pa(ː)-pa/. Shading indicates 95% confidence interval.

Figure 2:

Mean smoothed curves in hertz of the four tones produced by the male participants of the current study: first /a/ of /pa(ː)-pa/. Shading indicates 95% confidence interval.

1.4 The current study: perception of newly emerged and residual cues

Based on previous production data, the current study investigates the equilibrium of plosive voicing, phonation, and pitch height in the perception of high versus low tones in Risiangku Tamang. In early 2017, we conducted a tone identification experiment in Nepal with native speakers of Risiangku Tamang, adapting as much as possible a laboratory setting to the field, with the following questions in mind:

1. Pitch and phonation are used in signaling the tonal contrast in production. Are they equally important in perception, and how do they interact?

2. Plosive voicing is marginally produced and is highly redundant. What, then, is its role in perception, and how does it interact with pitch and phonation?

3. How do these features evolve after the phonemicization of the new tonal contrast? Will Risiangku Tamang necessarily follow the path to a more evolved variety, in which pitch becomes the primary cue? Can we observe evidence for such an evolution in apparent-time data?

2.1 Base words

There are almost no quadruplets with solely a tone difference. We selected a quasi-quadruplet relatively easy to synthesize: /¹ pa-pa/ ‘to be too liquid’, /² paː-pa/ ‘to be rough (taste)’, /³ pa-pa/ ‘to bring’, /⁴ paː-pa/ ‘to pile up’ (with numbers representing tones from highest to lowest). (The phonetic realization of the second /p/ in each word is voiced, as explained in note 2, Section 1.3.) Each word is a monomorphemic stem followed by an affix. They are common words, but /⁴ paː-pa/ ‘to pile up’ is less familiar to young speakers. All stimuli were synthesized with equalized segmental duration so as to minimize the effect of vowel length.

2.2 Stimuli

Stimuli were synthesized with target words preceded by a deictic /² tsu/ ‘this’, using an articulatory synthesizer, VocalTractLab 2.1 (Birkholz 2013), followed by other modifications detailed below. The target words were C₁ V₁ C₂ V₂ sequences, where C₁ and C₂ were both labial plosives, and V₁ and V₂ were /a/. Five parameters were manipulated on the target word. Examples of spectrograms are given in Appendix A.2 of Supplementary material. All stimuli can be found in the Supplementary material.^[3]

1. Degrees of breathiness of C ₁ V ₁: labeled here as “modal”, “breathy”, and “super breathy (sup_br)”. Three degrees of glottal opening were synthesized in VocalTractLab 2.1 to simulate three degrees of breathiness, by modifying the upper and lower rest displacement and the arytenoid area (see gestural scores in the Supplementary material). Their difference in breathiness was confirmed by H1–H2 measure over the target vowel: 2.5, 7, and 12 dB, for modal, breathy, and super breathy, respectively. Segmental durations and intensity contours were then equalized across items.

2. Presence or absence of the prevoicing of C ₁: VOT at −70 or 12 ms. For each unvoiced stimulus, we created a version with prevoiced C₁ with a voice lead of 70 ms computed as a low-pass filtered and attenuated extract of the vowel following a voiceless plosive.

The resulting sequences were then manipulated for F0 with TD-PSOLA (Valbret et al. 1992) implemented in Praat (Boersma and Weenink 2016) (F0 was at 115 Hz steadily on the vowel in /² tsu/):

1. F0 onset of V ₁: at 115, 130, 145, or 160 Hz.

2. F0 slope of V ₁: linear rise of 10 Hz, versus fall of 20 Hz.

3. F0 slope of V ₂: linear rise of 5 Hz, versus fall of 15 Hz. The F0 onset of V₂ is always 5 Hz lower than the F0 offset of V₁.

The endpoints of F0 onset were based on the production of speaker M3 in Mazaudon and Michaud (2008) (see Michaud and Mazaudon 2006: Figure 1), which are comparable to the F0 production in our data in Figure 2. The breathy and super breathy versions were created after trials of different parameters to meet the auditory judgment of breathiness of the first author, a trained phonetician. This made a total of 96 stimuli (3 degrees of breathiness × 2 voicing parameters × 4 F0 onsets × 2 F0 slopes of V1 × 2 F0 slopes of V2).

2.3 Participants and data collection

The results of 28 participants (14 female, 14 male) of the experiment, aged 33–79 (mean = 49, SD = 12.3) at the time of the experiment, will be reported.^[4] All are bilingual speakers of Risiangku Tamang and Nepali. We discarded the data of five additional participants who were unable to complete the task.

The experiment was conducted using PsychoPy 2 (Peirce and MacAskill 2018). Participants were tested individually in a recording studio in Kathmandu, or a quiet room in the village of Risiangku. They wore a Sennheiser HD518 headphone and responded on a laptop computer. In each trial, an auditory stimulus was presented in two repetitions, and the participant was invited to select the most appropriate among four pictures displayed in the four corners of the screen (see Appendix A.1 in Supplementary material). The response time-out was set to 10 s, in consideration of the unfamiliarity of the participants with this type of task. Participants took two short breaks during the experiment session. Stimuli were presented in a different randomized order for each participant.

The experiment session was preceded by two training sessions during which participants were presented with natural stimuli produced by two male speakers. The first training session of 11 trials aimed at familiarizing all the participants with the task, using different items from the target words. The second training session of 10 trials was designed to make sure that participants could identify each naturally produced target word from the quasi-quadruplet and associate it with the intended meaning represented by the picture. Feedback was given for correctness with a happy or a sad emoticon. Participants who had difficulties in understanding the instructions repeated the failed session once or twice. For most participants, we recorded their production of the four test words for qualitative observations.^[5]

2.4 Limitations

Since only one quadruplet was used, our results are limited to one place of articulation (bilabial) followed by the vowel /a/. It is possible that the perception of VOT and/or F0 is affected by place of articulation (Peralta 2018) and/or the following vowel. The reason why we refrained from including more words or repetitions was to avoid imposing a long attention span on our participants, who are mostly unaccustomed to computers and sometimes have difficulties understanding the instructions. As it was, most villager participants spent more than 30 min in completing all the sessions. Consequently, the limited number of trials may result in failure to detect an existing statistical significance due to a low statistical power (e.g., Kirby and Sonderegger 2018). However, we think that an increase in the number of trials could increase error rates, which would be undesirable. Another drawback of our study is that we were unable to establish a sociological profile for each participant. We will examine the age factor in the following, but other sociological factors were not controlled for.

3.1 Identification rates and response times

We first tried exploring how listeners identify the four tones. Possibly due to the equalized vowel length and the oversimplified F0 slope manipulation in our stimuli, the results were inconclusive, except that the ratio between tone 3 and tone 4 responses increased as the degree of breathiness increased (modal: 0.87; breathy: 1.11; super breathy: 1.46), in line with the production results in Mazaudon and Michaud (2008). Since we were mostly interested in the perceptual difference between high and low tones, in the following analyses, we group tones 1 and 2 into high tone responses, and tones 3 and 4 into low tone responses. Also, we do not expect F0 slopes of V1 or V2 to interfere in the identification of high versus low tones. Note that although each stimulus was presented only once to the participants, the averaging of responses over F0 slopes means that there were four tokens for each F0 + phonation + prevoicing condition.

Figure 3 shows the averaged identification rate of high tones. In the following, we provide descriptive statistics as well as the results from a generalized linear mixed model fitted to the response data (using the lmertest package [Kuznetsova et al. 2017] in R [R Core Team 2019]; formula and outputs in Appendix C.1 of Supplementary material). As expected, high tone identification rate increases as F0 onset increases (14% < 33% < 58% < 71%, β = 3.85 ,  SE = 0.18 , z = 21.28 , p < 0.0001 ), and as breathiness decreases (31% < 43% < 58%, β = − 1.68 ,  SE = 0.12 , z = − 13.73 , p < 0.0001 ). Unlike the marginality of prevoicing in the production of low tones, prevoicing leads to a strong bias towards low tone perception (24% for unvoiced vs. 64% for prevoiced stimuli, β = − 3.67 ,  SE = 0.34 , z = − 10.73 , p < 0.0001 ). The model also shows that age interacts with two main factors: the effect of prevoicing decreases with age ( β = 0.05 ,  SE = 0.03 , z = 2.05 , p < 0.05 ), while the effect of phonation increases with age ( β = − 0.03 ,  SE = 0.01 , z = − 3.68 , p < 0.0005 ).

Figure 3:

High tone identification curves (T1/T2) by F0 onset, phonation, and prevoicing. Error bars represent one standard error.

Two asymmetries can be observed. First, at the highest F0 onset, prevoiced stimuli reduce high tone responses almost by half, regardless of phonation, while unvoiced stimuli do not have the opposite effect at the lowest F0 onset. Second, for unvoiced stimuli, modal voice leads to a strong bias towards high tone identification, including at the lowest F0 onset, while breathy and super breathy stimuli do not bias toward low tone identification at the highest F0 onset. (See Appendix C.1 in Supplementary material for results of pairwise comparisons, using the emmeans package [Lenth 2019].) Thus, the prevoiced property conflicts with high tone identification, and modal voice with unvoiced plosive conflicts with low tone identification.

Response times are also affected by the (in)congruence of cues. Figure 4 shows that response time is the lowest when all three cues are congruent: (1) breathy voice + prevoiced + lowest F0, or (2) modal voice + unvoiced + highest F0. Response times increase when two cues are in conflict, where the identification rate is around 50%: (1) prevoiced + highest F0, or (2) modal voice + lowest F0. (See Appendix C.2 in Supplementary material for statistical models.)

Figure 4:

Response time curves by F0 onset, phonation, and prevoicing. Error bars represent one standard error.

3.2 Classification tree analysis

A classification tree analysis (CART) was conducted to further assess the relative importance of each cue in classifying high versus low tone categories, using the rpart package (Therneau et al. 2019) in R. Note that this data classification method divides the data into two maximally homogeneous groups at each split, but does not say anything about how a listener makes decisions about a tone category.

Figure 5 shows the classification tree, plotted by the rpart.plot package (Milborrow 2019) in R. F0 is used in the first split, with 138 Hz as cut-off. When F0 is above 138 Hz, prevoicing is the best separator between the two response categories. When F0 is below 138 Hz, the prevoiced stimuli predict a low tone response, and unvoiced stimuli with a breathy voice also predict a low tone response. The asymmetries observed in Figure 3 are reflected here again to some degree. First, across the board, prevoiced stimuli predict a low tone response, while unvoiced stimuli have less powerful predictability. Second, for unvoiced stimuli with lower F0, modal voice predicts a high tone response, whereas for higher F0, phonation plays a smaller role in the prediction of the tone response.

Figure 5:

Classification tree of high versus low tone response categories. Below each node, the percentage of observations of each response category in the node is given.

3.3 Apparent-time variation

For each factor, high tone response differential was calculated for each participant by subtracting the minimum value (i.e., responses to stimuli with lowest F0 onset, or prevoicing, or super breathy phonation) from the maximum value (i.e., responses to stimuli with highest F0 onset, or without prevoicing, or with modal phonation). Higher response differential is taken as an indication of a greater use of a given cue by a given participant (see, e.g., difference score used in Idemaru et al. 2012). As shown in Figure 6, age correlates negatively with the response differential based on prevoicing, and positively with the response differential based on phonation, in line with the results of the generalized linear mixed model reported in Section 3.1. This suggests an increase in the use of prevoicing and a decrease in the use of phonation with decreasing age (for further analysis, see Appendix C.3 in Supplementary material).

Figure 6:

Scatterplot of high tone response differentials against age with fitted regression lines and coefficients: left, for F0; center, for prevoicing; right, for phonation.

4.1 Breathy voice as unmarked

The strong conflict between modal voice and low tone identification raises a question: why should phonation matter more when associated with low F0 than with high F0? This has no obvious psychoacoustic motivation. No such asymmetry has been found in the perception of sine-wave overtones (Kuang and Liberman 2018: Figure 4). As a comparison, in a study on Shanghai Wu with similar stimuli settings, it has been found that the effect of phonation is the smallest at the lowest F0 onset (Gao et al. 2020: step 7 in Figure 5), but increases as F0 onset increases.

Breathy phonation is usually interpreted historically as a step on the way to the devoicing of initial voiced plosives. But in modern TGTM languages, breathy voice is present with low tones across the board, including on words with sonorant initials which were and remained voiced. In the case of sonorants – as opposed to plosives – the modified series was not the voiced series, but the voiceless series, which developed high pitch accompanied with modal phonation as they became voiced. This, coupled with the strong conflict between modal voice and low F0, suggests that a slightly breathy voice might be the neutral setting for Risiangku Tamang, while modal voice is “marked”. A detailed study of phonation after continuant initials active in a tonal split would be instructive.

4.2 Perception-production delinking: the fate of plosive voicing

We found that the weight of plosive voicing is greater in perception than in production in Risiangku Tamang. Does this constitute an argument in the debate on production-led versus perception-led sound change (for a review, see Schertz and Clare 2020)? We think not.

Previous studies on incipient tonogenesis (e.g., Coetzee et al. 2018 on Afrikaans) or obstruent devoicing (e.g., Pinget et al. 2020 on Dutch; Gao and Arai 2019 and Gao et al. 2019 on Japanese; and Brunelle et al. 2020 on Chru) have shown that plosive voicing remains an important cue in perception while being variably used in production. These studies may be taken as arguments against the strong version of Ohala’s (1981) hypothesis that misperception, taken in a broad sense, is the source of sound change.

However, the Tamang situation differs in several respects from languages undergoing incipient tonogenesis like Afrikaans (Coetzee et al. 2018) and Central Malagasy (Howe 2017). First, we are dealing with a tonal split, not the emergence of tone in a previously toneless language. The presence of tones in the protolanguage was plausibly a favoring context for the reinterpretation of coarticulation features issued from laryngeal characteristics of the initial consonants as tones. Second, the protolanguage had two series of continuants, unlike, for example, Afrikaans (Coetzee 2017). Other Tibeto-Burman languages that are undergoing tonogenesis – the creation of primary tones (e.g., Kurtöp: Michailovsky and Mazaudon 1994; Hyslop 2009; Peralta 2018) – or a tonal split (e.g., Lalo: Yang et al. 2015), although more similar to the Tamang situation in some respects, evidence a much lower degree of devoicing of their initial plosives.

Thus, Risiangku Tamang, with its much higher rate of devoicing, and situated after the phonemicization of the new tonal system due to the merger of two series of proto-continuants, is further advanced on the time scale of a change. Evidence from Risiangku Tamang does not bear on an incipient change but may be considered consistent with the reframing of Ohala’s hypothesis by Kuang and Liberman (2018) and Pinget et al. (2020) along a time dimension in which production leads perception in the later stages of a change.

In fact, the situation may no longer reflect an ongoing change: after the completion of the change, listeners still attend to residual features as long as those features participate in the phonological categorization. Although plosive voicing is only marginally produced in Risiangku Tamang, when it is present, it cues low tones exclusively (since it never co-occurs with high tones). In this sense, the perceptual reliability of plosive voicing is greater than that of the other two features of pitch and phonation, which vary on a continuous scale.

4.3 Apparent-time variation, evolution, and language contact

Our results suggest that, as age decreases, the contribution of phonation decreases, while that of plosive voicing increases. The perceptual down-weighting of phonation is in line with the direction of the tonal split process. We would expect the same trend for plosive voicing, but our results suggest the opposite direction. One possible explanation is the higher proficiency of younger Tamang speakers in Nepali, in which plosives have a phonological voicing contrast. Tamang villagers of 40 years ago did not distinguish /b, d, g/ from /p, t, k/ when speaking Nepali. The saliency of the voicing feature in Nepali, acquired in their second language, was likely transferred by younger speakers to their perception of prevoicing in their native language. If this is true, it would suggest that a contact-induced sound change, independent from the tonal split, is taking place in parallel (see Pearce [2009] for a similar sociolinguistic situation, due to contact with French, among town-dwelling male speakers of Kera). A comprehensive sociolinguistic study would be needed to tease apart language-internal and contact-induced factors.

4.4 Concluding remarks

In conclusion, we have shown that in Risiangku Tamang, newly emerged and residual cues from the tonal split process are all used in perception and enter in interaction with other cues in a stable variation. This situation is frequently encountered in voice register languages. Our data shows that it also exists in a tonal language after the establishment of a new tonal contrast. Whether or not Risiangku Tamang eventually evolves towards the disappearance of all residual cues, its present transitional state can last a long time and be made more complex by language-external factors such as contact.