2.1. Dàgáárè background

Dàgáárè is a Mabia (formerly known as Gur; Bodomo, 2017, 2020) language of the Niger-Congo phylum with about 1.5 million speakers in Ghana and Burkina Faso (Angsongna & Akinbo, 2022). Dàgáárè is predominantly spoken in northwestern Ghana and southern Burkina Faso. There are four broad dialects of Dàgáárè, namely Northern Dàgáárè [dàgàrà], Southern Dàgáárè [wá:lɪ́], Western Dàgáárè [bɪ̀rɪ̀fɔ̀], and Central Dàgáárè [dàɡáárɪ̀]. Previous impressionistic fieldwork research suggests that Southern, Western, and Northern dialects have nine vowels (Bodomo, 1997; Kuubezelle & Akanlig-Pare, 2017; Ali, Grimm, & Bodomo, 2021), but recent acoustic and articulatory studies show that the Central dialect has a tenth vowel /ə/, which is the ATR counterpart of /a/ (Ozburn, Akinbo, Angsongna, Schellenberg, & Pulleyblank, 2018; Lloy, Akinbo, Angsonga, & Pulleyblank, 2019; Angsongna & Akinbo, 2022). The vowel inventory is shown in Table 1, with the ATR low vowel in parentheses since it may not exist in all dialects.

Dàgááre vowels can be grouped into ATR and RTR classes, as shown in Table 1. Like all Mabia languages, Dàgáárè has ATR harmony: Vowels within a word obligatorily agree in their ATR/RTR feature. Within words and in clitic groups in the language (i.e., with a single lexical item and any suffixes or clitics dependent on it), vowels must be harmonic. In this case, the root vowels trigger ATR harmony, targeting all vowels of affixes or clitics. Consequently, the vowels of affixes and clitics are either ATR or RTR depending on the root vowel. Some examples are shown in Table 2.

All vowels undergo ATR harmony in the Central dialect (Angsongna & Akinbo, 2022), but there is debate about the participation of the central vowel /a/ in ATR harmony in the nine-vowel dialects (Bodomo, 1997; Kuubezelle & Akanlig-Pare, 2017; Ali et al., 2021). As such, we exclude low vowels from the present study, so that this issue has no bearing on our experiments.

Disharmonic vowel combinations are impossible within a word, with the exception of compounds, which can involve ATR+RTR or RTR+ATR combinations. Consider the examples of ATR+RTR compounds below in Table 3, representing examples from all varieties of the language. Given that such disharmonic compound words are possible, we in fact expect speakers to be able to adequately distinguish disharmonic sequences, as they are exposed to them in the language.

Disharmonic compounds in the opposite order, with RTR followed by ATR noun roots, are attested in all dialects except in Dagara. In the Dagara variety, the ATR roots in a compound can trigger harmony which targets the preceding RTR roots, as shown in Table 4. In contrast, no harmony happens in ATR+RTR compounds like those in Table 3, even in Dagara. Consequently, tongue-root harmony is considered ATR-dominant and regressive in varieties like Dagara (Casali, 2003).

Overall, harmony in Dàgááre is root-controlled, making it difficult to determine dominance and directionality directly. The Dagara dialect is also a root-controlled system, but with its harmonizing of compounds, it shows an underlying ATR-dominant and regressive nature to the harmony. Other dialects, particularly Central Dàgááre, do not harmonize compounds at all, which is common in many harmony systems. It would be incautious to make a strong statement about the whole language based on evidence from a single dialect. That said, we note that nothing in Central Dàgááre is inconsistent with underlying ATR dominance and regressive directionality. We cautiously propose that all varieties of Dàgááre have this basic nature, but the evidence only appears in Dagara, because it is the only variety known to harmonize in compounds. However, we note that further investigation into and characterization of the underlying nature (i.e., where root-control is irrelevant) of this harmony system is still needed, and we leave that to future research.

2.2. Perception of (vowel) harmony background

Claims about how harmony is motivated by perceptibility are relatively common in the literature (e.g., Kaun, 1995; Ohala, 1994), but there has been very little perceptual work investigating these hypotheses directly. In this section, we overview some of the research on these hypotheses.

In Finley’s (2012) study, English speakers were tasked with learning a rounding harmony pattern in an artificial language; while this study is not directly a perception task, it relates to perceptibility claims because the artificial pattern had high vowel harmony triggers in one condition and mid vowel harmony triggers in the other. The latter pattern is the more common one cross-linguistically: Kaun (1995) posits that, rather than being the mere result of something like historical change or language relatedness, this asymmetry exists because (1) mid vowels are less likely to be perceived and produced as canonically round, and (2) rounding harmony enhances perception of the [round] feature by repeating it across the domain (i.e., word). Participants were found to perform better (i.e., they more accurately learned the pattern and correctly generalized it more often to novel instances) in the mid trigger condition, implying that the bias towards perceptual enhancement exists in the learning of harmony patterns.

Kimper (2017) also conducted multiple experiments wherein English-speaking participants listened to a nonce word and then had to report whether a particular target vowel was present in that word. The nonce word always contained a transparent vowel in some position, while its other vowels were either harmonic or disharmonic for height and backness with each other and with the target vowel. Participants responded more quickly and more accurately when the target vowel was harmonic with the nonce word’s vowels along some dimension. Again, this result implies that harmony facilitates perception of particular vowel features, though the effect was attenuated when the transparent vowel interceded between any of the harmonic vowels.

Similarly, Kimper (2011) conducted an AX task with English speakers, using the feature ATR crossed with a height feature. He hypothesized that ATR harmony maximizes discriminability, leading to faster and more accurate responses when both vowels in a disyllabic stimulus are distinct than when only one. He also hypothesized that this advantage would be reduced when there is disagreement in height between the vowels. Both of his hypotheses were supported by his experimental results.

Other studies have returned results that seem to reject the aforementioned hypotheses. Grosvald (2009) investigated the perception and production of anticipatory vowel-vowel (VV) ([i/a]-[ə]) coarticulation by English speakers. While some participants did consistently show coarticulation on the target (ə) in their productions, and correctly perceived [i]-/[a]-colouring on the [ə] in others’ productions, the within-subject correlation between production and perception was positive but weak. These findings oppose Ohala’s (1994) theory that vowel harmony arises from listeners’ perception of VV coarticulation becoming grammaticalized over time. In another study looking at VV coarticulation (Busà & Ohala, 1999), American English and Italian speakers listened to V1-C-V2 sequences and C-V(2) sequences, where V2 was on a continuum between [i] and [u], and V1 (if present) was always [i]. Participants perceived V2 as being more back in the disyllabic context (i.e., preceded by [i]) even when there was a long interval between V1 and V2. That this dissimilatory effect persists even over long periods indicates that, rather than stemming from compensation for perceived coarticulation carrying over from the V1 (which could reasonably only happen in fluid speech between vowels that are temporally close to each other), this is the result of more general contrast effects in perception. Again, we find opposition to Ohala’s (1994) theory: If it is the case that contrast effects (basically listener hypercorrection) are at play when one listens to VV sequences, then how could vowel harmony ever arise if it is supposedly the result of listener hypocorrection? Some additional contrary evidence is also found in electroencephalography (EEG) studies. In these studies, event-related potentials (ERPs) between 50–240ms after stimulus onset are thought to correspond to activation of phoneme representations for native speech (see Gansonre, Højlund, Leminen, Bailey, & Shtyrov, 2018 for a review). If vowel harmony aids the live perception of speech sounds, then vowel harmony ought to be associated with ERPs at these early time intervals; however, some EEG studies have found that vowel harmony is only associated with later-occurring ERP components (if any). For example, Tuomainen (2001) conducted a Finnish auditory segmentation task and found that vowel harmony was associated with an N400 (negative deflection 400ms after stimulus onset, which is thought to reflect higher-order types of language processing such as word recognition and integration [Junge, Boumeester, Mills, Paul, & Cosper, 2021]). This result indicates that vowel harmony (and disharmony) could help listeners correctly and more quickly identify boundaries between words (e.g., a speaker of an ATR harmony language would know immediately to separate [begedɪgɪ] into two harmonic units, [bege] and [dɪgɪ]). However, the result also indicates that vowel harmony does not help listeners identify individual sounds, in contrast to the aforementioned hypotheses. In brief, direct experimental investigations into the effects of vowel harmony on speech perception are few, they are often conducted with speakers of non-harmony languages, and they have returned largely mixed results.

Turning now to how these proposals might be applicable to ATR, there are proposals specifically suggesting that vowel harmony aids the perception of ATR contrasts, particularly of ATR contrasts that are considered ‘difficult’, such as those on high vowels. These proposals have, for the most part, not been tested experimentally. However, we can consider what the proposals of harmony more generally would predict when applied to ATR harmony. Considering Kaun’s (1995) claim that rounding harmony extends perceptibility of a feature, spreading from vowels where rounding contrasts are weakly perceived to those where the perceptibility is strong, we might expect ATR harmony to similarly be triggered by sounds on which ATR is difficult to perceive, and target those where the contrast is most difficult. However, it is unknown experimentally which ATR distinctions are the most difficult to perceive, as evidence on perception of ATR is typically anecdotal. Our experiments will contribute to filling this gap. Interestingly, while the rounding harmony that Kaun discusses tends to be parasitic on height (e.g., in the Yowlumne dialect of Yokuts, rounding harmony only applies when the trigger and target vowels are of the same height [Kuroda, 1967]), ATR harmony often behaves in a different way: It is not usually parasitic, but instead it is common for a certain class of vowels to be ‘dominant’ (Casali, 2008). This means that it is difficult to apply Kaun’s discussion of instances where rounding contrasts are difficult to perceive directly to thinking about ATR. For instance, there are ATR systems where /i/ triggers harmony but /ɪ/ does not, because [+ATR] is the dominant feature (see Rose, 2018 for a review); presumably, if the contrast between these two vowels is difficult to perceive, it would be of perceptual benefit for both to trigger harmony, rather than just one. In terms of Ohala’s hypocorrection hypothesis, we generally expect listeners to attribute coarticulatory ATR effects to the target vowel. Coarticulation, like ATR harmony, tends to be anticipatory (e.g., Hyman, 2002), which means that we might expect the initial vowels in disyllabic forms to be subject to hypocorrection. However, little is known specifically about ATR coarticulation, so the question of which types of segments we would expect to see as most perceptually susceptible to hypocorrection is a difficult one. Overall, then, linking these previous claims to ATR harmony is complex given substantial gaps in the current understanding of ATR, and so this paper aims to establish a baseline for perception of ATR such that some of these more complex issues can be addressed in the future.

In the most general case, however, disregarding triggers and targets, the expectation from these previous works is clear: If the perceptibility hypothesis holds, then when asked to distinguish two similar words, we expect participants to perform better with the assistance of harmony. For instance, distinguishing ɪ-ɪ sequences from e-e should be easier than distinguishing either of those from e-ɪ sequences: There are two opportunities to hear the ATR feature in the harmonic forms, but only one in the disharmonic forms. Such a result would be in line with Kimper’s (2011) results with English-speaking participants.

2.3. Our study

To investigate this and related hypotheses, two ABX discrimination task experiments were conducted. The ABX task was selected because it is a popular paradigm in perceptual experiments, and because of its ease of explanation and presentation. In the first experiment, participants were tested on their ability to distinguish three vowel contrast types in monosyllabic (CV) and harmonic disyllabic (CVCV) nonce word contexts. This experiment served as a baseline to establish listeners’ ability to discriminate the vowel contexts and to examine whether harmony helps discrimination. In the second experiment, participants were tested on a single contrast type in monosyllabic (CV) and both harmonic and disharmonic disyllabic (CVCV) nonce word contexts. This experiment served to examine whether disharmony impedes perception relative to harmonic and monosyllabic contexts. The next section presents more details about Experiment 1, followed by Section 4 which deals with Experiment 2.

3.1. Methodology

3.1.2. Materials

Nonce words of form CV and CVCV were produced by a phonetically trained male Dàgáárè native speaker. To increase comparability across conditions, the only consonant used was [k] (always as an onset), and the only tone used was low. We chose a stop onset to reduce interactions with surrounding vowels, and we chose a velar in order to avoid real words ([ti], [di], [pi], and [be] with low tone are all words in Dàgáárè). The voiceless velar stop was chosen over the voiced option to increase the sonority differential between the onset and the vowels of interest.

Given that there are no known differences in harmony operation by backness in the language (i.e., both front and back vowels participate in harmony in the same way), we used only front vowels in this study to decrease the length of the experiment, while increasing the number of repetitions per A/B pair in the ABX tasks.

For Experiment 1, we used four nonce words in the form CV (i.e., [ke, kɛ, ki, kɪ]) and eight in the form CVCV (i.e., [keke, kɛkɛ, kiki, kɪkɪ, keki, kike, kɛkɪ, kɪkɛ]). All nonce words were harmonic, making them possible but non-existent Dàgáárè words.

Stimuli were recorded by a phonetically trained male native speaker of Dàgáárè, using a Shure WH30 headset microphone at the sampling rate of 48.1 kHz in WAV format. The microphone was connected to a Zoom Q8 video recorder. The audio from the microphone was saved as a separate file at the same time as the video file. Six to eight tokens of each nonce word were recorded. The best recordings were selected, determined based on representativeness of the intended category as judged by the authors and the phonetically trained native speaker. Individual trials never included identical tokens. For all selected tokens, all authors and the phonetically trained native speaker agreed that the IPA transcription matched the intended nonce word. Stimuli were segmented in Praat (Boersma & Weenink, 2022).

To ensure that the vowels in the stimuli were accurately produced, we plotted them prior to running the perception experiment. Formants were measured by a Praat script at the midpoint of each vowel, with the vowel boundaries determined as where clear periodicity began and ended. This includes all vowels in both experiments and in both positions, and it was particularly important given that some of the stimuli in Experiment 2 were illicit as (non-compound) Dàgáárè words. The formant plot is shown in Figure 1, and a table of formant means and standard deviations is shown in Table 5. A graph of the vowel durations is provided in Figure 2. All graphs and statistics in this paper were done using R (R Code Team, 2022) in RStudio (RStudio Team, 2022).

Figure 1

Formant plot for front vowels in Dàgáárè.

Table 5

Mean and Standard Deviation of formant values and duration.

		i	ɪ	e	ɛ
F1	Mean	247.90	330.58	308.73	440.01
	SD	22.42	43.59	19.70	31.63
F2	Mean	2095.11	1858.47	1983.77	1807.46
	SD	112.65	172.91	157.62	102.27
Duration	Mean	94.05	84.40	107.73	130.42
	SD	23.17	20.35	15.38	18.50

Figure 2

Vowel durations in the stimuli.

3.2. Hypotheses

Experiment 1 consisted of three vowel contrast types and four contexts. As noted, ATR contrasts in high vowels are predicted to be particularly difficult to perceive based on typological findings (Rose, 2018), and high RTR/mid ATR contrasts (e.g., e/ɪ) have been found to be difficult in previous experiments on other languages (Rose et al., 2023). Additionally, high RTR/mid ATR vowels are acoustically overlapping in Dàgáárè (see Figure 1 in the Materials section). As such, we predicted that e/ɛ contrasts will have the highest accuracy, with significantly lower accuracy in the other contrast types. Given that Rose et al. (2023) found that acoustic overlap was the greatest predictor of perception results, we predict that the cross-height e/ɪ contrast will have the lowest accuracy rates.

In terms of contexts, harmony is said to aid in perceptibility, and all our disyllabic contexts were harmonic, giving participants more opportunities to hear the harmonic feature: They have two opportunities to hear ATR/RTR instead of just one, because both vowels differ. This prediction was discussed in Section 2.2 as the most general case that we would expect based on perception-related hypotheses in the harmony literature and based on previous experiments like Kimper (2011). As such, we predict that all disyllabic contexts should have higher accuracy than the monosyllabic context. Within the disyllabic harmonic contexts, we noted above that ‘context’ is effectively the same as ‘contrast’, just in the V2 position rather than V1. As such, we predict that accuracy should reflect the same contrast-specific considerations noted above. Specifically, since mid vowels are predicted to be the easiest to accurately distinguish, in part due to the lack of acoustic overlap of ɛ with any of the other vowels, the harmonic mid context (e/ɛ) should have the greatest accuracy, followed by harmonic high (i/ɪ), followed by cross-height (e/ɪ), in the same way as for the contrast types.

The hypothesis follows because mid vowels are expected to have the strongest perceptual contrast, they should also be the best place to hear the effects of harmony in that ɛ is quite distinct from the other vowels in the language, and so they should improve performance the most in harmonic disyllables.

3.3. Results

The proportion correct for each participant for each trial type was calculated. Null responses (i.e., trials where a participant did not respond) were excluded from this calculation; this decision resulted in the exclusion of one participant in Experiment 1 whose responses all registered as null. As such, the results are based on 23 participants in Experiment 1. Of the participants who were included, a total of 184 trials (average of eight per participant) were excluded for being null responses. We then calculated the mean and standard deviation of reaction time for each participant, for all trials with non-null response, and excluded any trials where a participant’s reaction time was greater than 2.5 standard deviations away from their mean. This resulted in the exclusion of an additional 125 trials total (average of 5.4 per participant).

Responses were submitted to a linear regression mixed effects models on proportion correct (Baayen, Davidson, & Bates, 2008; Jaeger, 2008). All statistical models throughout the paper were implemented using the function lmer from the package lme4 (Bates, Mächler, Bolker, & Walker, 2015). Proportion correct was calculated for each subject for each condition (i.e., for each subject in each context/contrast combination). The range in total numbers of tokens that went into each proportion correct was 10 to 16, depending on how many null responses and responses excluded due to reaction time a participant had for a given condition. The mean number of included tokens in each proportion correct for Experiment 1 was 14.88 and the standard deviation was 1.28.

For all models, the random effect was a by-participant random intercept; item effects were not included as random intercepts due to collinearity with the fixed effects, and the models did not converge with random slopes. It is worth noting that the Participants random effect, along with the fact that our conclusions are based on differences between stimulus pairs, means that any extra by-subject variability due to online presentation should not have affected the results.

The fixed effects were Context and Contrast. Context has levels i/ɪ, e/ɪ, e/ɛ, and monosyllabic (reference level); Contrast has levels i/ɪ, e/ɪ, e/ɛ (reference level). The default dummy coding scheme for categorical variables was used. We also ran a version of the Experiment 1 model for comparison purposes that included the interaction between context and contrast. Using a model comparison ANOVA, we determined that the BIC of the model without the interaction term was lower than the model with the interaction (BIC –274.87 vs. –254.92). As such, the model without interactions performed better, and is therefore the one reported here.

Note that we did not consider whether X was A or B as an effect, because the experiment was counterbalanced by this property, and so any effect of this nature (e.g., if any participants were naturally biased towards choosing A over B) would be balanced over the course of the experiment and would not affect the results.

The reference levels in the model were chosen with our hypotheses in mind. We hypothesized that the e/ɛ contrast will have significantly better performance than other contrasts, due to the acoustic separation. The model, therefore, tests the other contrasts in comparison to this one, and we expect performance will be worse on i/ɪ and e/ɪ compared to e/ɛ. In terms of contexts, we hypothesized that the three disyllabic contexts, all of which are harmonic, will show better performance than monosyllabic contexts, due to opportunities to hear the ATR feature on both vowels. The model compares the three disyllabic contexts against the monosyllabic one; we expect performance to be better on the disyllabic contexts. However, in order to test additional nuances, we also re-levelled the factors and ran the model with other reference levels. Doing so allows us to see whether the hypothesized accuracy trend of e/ɛ > i/ɪ > e/ɪ holds in the data, as not all of these comparisons can be directly tested with a single reference level. This method allows us to compare other levels of the factors to test our hypotheses in greater detail.

The monosyllabic context and e/ɛ context had approximately equal performance (mean = 69.9% and 70.1% respectively), while other contexts had worse performance (means 66.8% and 64.8%), though only the “cross-height” (e/ɪ) context was significantly different from monosyllabic (p < 0.05; see Table 8). Moreover, both other contrasts (e/ɪ and i/ɪ) had worse performance than the mid e/ɛ contrast (means = 70.1% vs. 65.7%, 68.0%), though again the difference only reached significance for the “cross-height” [ɪ]~[e] contrast (see Table 8). Table 7 illustrates the mean and standard deviation of proportion correct for all context/contrast combinations in Experiment 1 (obtained using the function ddply from the package plyr (Wickam, 2011)), while Tables 8, 9, 10 show the model output. In Table 8, the intercept represents the level at the reference level, which is the contrast e~ɛ in the monosyllabic context, while the remaining rows compare particular context and contrast levels to those reference levels. In order to compare all levels directly, we also repeated the model with all other possible reference levels; Table 9 shows e~ɪ as the reference level for both contrast and context, while Table 10 shows i~ɪ as the reference level for both contrast and context. From these tables, we can conclude that Context e~ɪ significantly differs from the monosyllabic Context (p < 0.05) and e/ɛ Context (p < 0.05), Contrast e~ɪ significantly differs from the Contrast e/ɛ (p < 0.05), but no other differences are significant.

Figures 3 and 4 illustrate graphically the proportion correct by Context (faceted by Contrast and without facets respectively) and Figures 5 and 6 illustrate proportion correct by Contrast (faceted by Context and without facets respectively). All results graphs (excluding the acoustic plots in Section 3.1.2) were created using the function ggplot in the package ggplot2 (Wickam, 2016). The facets sort the data into separate graphs by the faceted criterion, while the non-faceted versions show all data combined. For example, Figure 3 shows three smaller graphs, one for each Contrast, and within each plot, the x-axis shows the Context and the y-axis shows the proportion correct, specifically for the given Contrast within that graph. Figure 4 has the same information, but not divided by Contrast, so that all three Contrast levels are combined into a single graph, again with Context on the x-axis and proportion correct on the y-axis. The violin plots show the density at a particular point, such that a point on the y-axis where the ‘violin’ is wider, like just above the 0.8 line for the e/ɛ contrast in Figure 4, represents a proportion correct that was very common in the results. Similarly, narrow parts of the violin plot show places where the density of respondents with that proportion correct was lower. Inside each violin is a standard boxplot, where the box shows the location of the first to third quartile of the data, with the thick line in the middle showing the median. The ‘whiskers’ of the boxplot, which are the lines going up and down from the box, show the range of the data minus any outliers.

Figure 3

Proportion correct by Context (=V2) (facet by Contrast (=V1)).

Figure 4

Proportion correct by Context (=V2).

Figure 5

Proportion correct by Contrast (=V1) (facet by Context (=V2)).

Figure 6

Proportion correct by Contrast (=V1).

To visually check that the results were not due to a single participant with particularly low or high results, we also created a boxplot of results by-participant, shown in Figure 7. In this graph, we see that we have a combination of participants who do quite well (e.g., participants 2, 6, 7, etc.) and those whose results are lower (e.g., participants 1, 5, 8, etc.). There are no participants who are visual outliers in the results, in that none of the participants are consistently getting zero correct and we have a general spread of participants across the accuracy levels. As such, it does not appear to be necessary to remove any participants for being extreme outliers to the results.

Figure 7

Proportion correct by Context for each participant.

To summarize these results, Experiment 1 found that participants had significantly lower accuracy on e/ɪ as both V1 (Contrast) and V2 (Context), compared to the monosyllabic and e/ɛ levels. The pairs e/ɛ and i/ɪ did not significantly differ from each other as contrasts, nor did they differ significantly from monosyllabic when looking at contexts. However, there was a trend for i/ɪ to have lower accuracy as both Contrast and Context compared to e/ɛ and monosyllabic.

In terms of the hypotheses, the trend matched what we expected within disyllabic words: Accuracy was highest for e/ɛ as Contrast/Context, followed by i/ɪ, followed by e/ɪ. Only e/ɪ was significantly different from e/ɛ. However, participants did not have greater accuracy on harmonic disyllabic than on monosyllabic words, in contrast to our prediction. Indeed, accuracy on the context e/ɪ was significantly lower than on monosyllabic contexts.

Experiment 1 suggests that e/ɪ is the hardest contrast for speakers. However, because comparisons were always between words with two differences in the disyllabic items, we cannot draw firm conclusions about the effect of harmony on perception. In contrast to our predictions, harmony does not aid perception, in that monosyllables are, on average, discriminated better than harmonic disyllables. However, Experiment 1 leaves open the possibility that harmony helps within disyllables. In other words, disyllables may be easier to discriminate when harmonic than when disharmonic. Experiment 2 addresses this possibility; it focuses on the e/ɪ contrast and includes disharmonic disyllabic items, in addition to monosyllabic and harmonic ones.

4.1. Methodology

4.2. Hypotheses

Experiment 2 had monosyllabic contexts (kɪ vs. ke), harmonic disyllabic contexts (kɪkɪ vs. keke), and four types of disharmonic disyllabic contexts (before ɪ, before e, after ɪ, after e), all using the vowel contrast that was the most difficult for listeners in Experiment 1. We predict that, compared to monosyllabic contexts, accuracy should be greater in harmonic contexts and lower in disharmonic contexts, given that harmony is said to aid perceptibility. Specifically, listeners might hypocorrect the disharmony in perception, such that we expect listeners to have the most difficulty with the four contexts comparing a disharmonic disyllabic nonce word to a harmonic word that differs in only a single vowel. Beyond hypocorrection, this result would also be predicted by the discussion in Sections 2.2 and 3.2 about one versus two differences: The harmonic context has a pair of words where both vowels differ, giving two opportunities to hear ATR, while the two nonce words in each disharmonic context differ in only a single vowel, resulting in fewer opportunities for listeners to be certain of the ATR features in each word.

Since harmony in Dàgáárè is root-controlled and bidirectional, we do not necessarily predict any directionality effects. However, at least in the Dagara variety of Dàgáárè, ATR is posited to be the dominant feature (Casali, 2003). Further, the harmony is said to be regressive (again, at least in Dagara), as noted with discussion of compounds in Section 2.1. Due to the apparent ATR dominance, we might expect [e] to show greater pressure to trigger harmony compared to [ɪ], because it is ATR. Under a theory where harmony is phonologized perceptual compensation, we might expect the degree of perceptual compensation to reflect properties like ATR dominance and regressivity. If so, then we might predict that disharmonic forms involving [e] in V2 position would be more likely to be inaccurately perceived as harmonic. Such contexts should then be more difficult and therefore show lower accuracy. In other words, if (a) [e] is more likely to trigger harmony, because it is ATR, (b) regressive is a more likely directionality, and (c) these facts are due to perceptual hypocorrection, then kɪke vs. keke should be particularly difficult to distinguish. Indeed, we expect listeners to incorrectly harmonize the former in perception.

4.3. Results

As for Experiment 1, the proportion correct for each participant for each trial type was calculated. Null responses (i.e., trials where a participant did not respond) were excluded from this calculation, resulting in the exclusion of a total of 86 trials with null responses (average of 3.4 per participant) in Experiment 2. As for Experiment 1, we excluded any trials where a participant’s reaction time was greater than 2.5 standard deviations away from their mean. This resulted in the exclusion of an additional 56 trials total (average of 2.2 per participant) in Experiment 2.

Responses were submitted to a linear regression mixed effects models on proportion correct (Baayen et al., 2008; Jaeger, 2008). As for Experiment 1, the random effect was a by-participant random intercept; item effects were not included as random intercepts due to collinearity with the fixed effects, and the models did not converge with random slopes. As before, it is worth noting that the Participants random effect, along with the fact that our conclusions are based on differences between conditions, means that any extra by-subject variability due to online presentation should not have affected the results. For Experiment 2, the fixed effect was Context, as there was only one contrast, as shown in Table 11. The levels were before e, after e, before ɪ, after ɪ, harmonic, and monosyllabic. Given that our experiment involved comparing monosyllabic contexts to disyllabic contexts and comparing harmonic contexts to disharmonic contexts, we ran the model twice, once with the monosyllabic context as the reference level and once with the harmonic context as the reference level. The former allows us to compare performance on monosyllabic contexts to all other (i.e., disyllabic) contexts, while the latter allows us to directly compare harmonic to disharmonic disyllabic contexts.

In Experiment 2, like in Experiment 1, all other contexts were again worse than the monosyllabic context, in terms of lower proportion correct, as shown in Table 12, though only slightly so for the harmonic context (means 68% vs. 67%). Using the model described above and illustrated in Tables 13, 14, this trend was significant (p < 0.05) for the contexts before and after [e], and for the context before [ɪ]. In other words, kɪke vs. keke, kekɪ vs. keke, and kɪkɪ vs. kekɪ were significantly harder to distinguish than kɪ vs. ke. The difference from monosyllabic was not significant for the context after [ɪ] (kɪkɪ vs. kɪke) or for the harmonic context (kɪkɪ vs. keke). Table 12 shows the means and standard deviations of proportion correct for each Context in Experiment 2, while Tables 13, 14 show the model output, comparing to monosyllabic and harmonic reference levels respectively. Figure 8 illustrates the results graphically. In Figure 8, the x-axis is Context and the y-axis is Proportion Correct, parallel to the graphs in Experiment 1, and the graph in Figure 8 is similarly a violin plot with a boxplot overlaid. Given that there was only one fixed effect, no faceted graphs were created. The graph shows clearly how harmonic and monosyllabic contexts are comparable (the violins and boxes are in similar places with similar densities), but the disharmonic ones are generally situated at a lower proportion correct, with the density (wider part of the plot) at lower proportion correct values. This graph therefore visually reflects the means and statistical results described earlier.

Figure 8

Proportion correct by Context for Experiment 2.

Similarly to Experiment 1, we created a boxplot of results by participant, shown in Figure 9. Again, there is a spread of results, but it does not appear that any participant is a particular outlier who might be unduly affecting results. It is worth noting that there are a few participants in Figure 9 who have higher accuracy on harmonic items than on monosyllabic ones, namely participants 4, 6, 10, 14, 20, 21, and 23.³ This pattern was the predicted one, but is not the majority one. We point it out because it suggests that individual variation may be an interesting direction to explore in future work.

Figure 9

Proportion correct by Context for each participant.

To summarize, Experiment 2 showed that most of the disharmonic contexts have significantly lower accuracy than the harmonic and monosyllabic contexts, indicating that participants did worse at distinguishing disyllabic disharmonic stimuli than at distinguishing monosyllables or harmonic disyllables. These results are shown directly in the statistical comparisons in Tables 13 (for the comparison to monosyllabic) and 14 (for the comparison to harmonic disyllables).

5.1. Summary of results

In terms of both contrast and context, only the “cross-height” (ɪ vs. e) factor was significant in Experiment 1. In other words, all trials involving [ɪ] versus [e] were more difficult than those with the other vowel pairs, with significantly lower accuracy rates. This result is expected, as the formant plot showed more overlap for [ɪ] versus [e] than for any other vowel pairs, and previous studies (e.g., Rose et al., 2023) have also found cross-height pairs (mid ATR with high RTR) difficult to differentiate in other languages. This fact motivated the choice of [ɪ] versus [e] as the contrast examined in Experiment 2.

In both experiments, all disyllabic contexts were harder for participants than monosyllabic contexts, in terms of lower accuracy rates, though in most (harmonic) cases not significantly so. This result may have simply been an effect of memory, as participants heard twice as many syllables in disyllabic than in monosyllabic trials. Specifically, participants may have had greater difficulty remembering what A and B were when it came time to listen to C, simply because they had already heard more syllables. It is worth noting that the inter-stimulus interval (500ms) was relatively long, meaning listeners needed to hold A in their memory for 1.5 seconds before responding. This long interval is likely to force participants to use phonological representations instead of auditory impressions, though we note that because participants hear competitors aloud in an ABX task, unlike in real-world language perception, it remains possible that they use auditory information in this task in a way that they do not in real-world phonological perception.

Notably, in contrast to Rose et al. (2023), accuracy rates were low across-the-board in these experiments, suggesting that these contrasts are generally difficult to distinguish even for (at least some) native speakers. The individual speaker graphs in Figures 7 and 9 show that some participants in our experiments did have very high accuracy, but there was a wide range in accuracy levels between participants, and on average the accuracy was low in our experiments. Even for the [e] versus [ɛ] contrast, which has no acoustic overlap according to Figure 1, participants in Experiment 1 showed less than 75% accuracy. We can therefore conclude from Experiment 1 that, even in cases without overlap, ATR is a difficult property to distinguish in Dàgáárè. Nonetheless, the comparison to the Rose et al. (2023) experiment should be made with caution, because of the differences in experiment design: That experiment was an AX task, while ours was ABX. Furthermore, their experiment was in-person before the pandemic, while ours was online during the pandemic. We return to this point later in the discussion.

With the caveat that memory is known to play a role in ABX tasks (Gerrits & Schouten, 2004), harmony does not appear to help perception of these difficult contrasts, given that performance was worse (though generally not significantly so) in harmonic disyllabic contexts than in monosyllabic ones. That said, disharmony does appear to hurt perception, with three of the four disharmonic contexts in Experiment 2 having significantly lower accuracy compared to monosyllables and harmonic disyllables.

Interestingly, the harmonic disyllables in Experiment 2 were [kɪkɪ] versus [keke], which was also the pair for the e/ɪ contrast in the e/ɪ context for Experiment 1. However, this pair had very different accuracy in the two experiments, as we can see in Table 7 (for Experiment 1) versus Table 10 (for Experiment 2). This pair had by far the lowest accuracy in Experiment 1, at 57%, whereas in Experiment 2, it had a much higher accuracy, at 67%, which is more comparable to the results from Experiment 1 for other harmonic pairs and even some of the monosyllabic pairs. We will return to this issue in the general discussion in Section 5.3.

5.2. Returning to the hypotheses

5.3. General discussion

As noted in the introduction, in-prep articulatory work on Dàgáárè (Akinbo et al., in prep) has suggested that ATR harmony may be challenging to motivate articulatorily in the language. As such, we have reason to expect listener-based motivation for ATR harmony, namely that there should be some sort of perceptual motivation for harmony to occur. This was not apparent in Experiment 1, given that disyllabic harmonic forms did not improve accuracy over monosyllabic forms. In other words, having two opportunities to hear the ATR value (in harmonic disyllables) did not improve accuracy over having only one such opportunity (in monosyllables). Moreover, in Experiment 1, we were particularly interested in testing the hypothesis following Kaun (2004) that certain vowels are good triggers or targets based on perceptibility of the contrast. For example, since e~ɛ are the most perceptibly distinct, we would expect kɪkɛ vs. keke to be significantly better than kɪ vs. ke, due to the additional e~ɛ contrast where the ATR distinction can be heard. Note that this issue does not directly require us to know which vowels are triggers or targets in the disyllabic forms, but simply relates to the fact that disyllabic forms provide two opportunities for hearing the ATR contrast, compared to a single opportunity in the monosyllabic forms. Additionally, e~ɛ is expected to be the easier place to hear the contrast given the acoustic distinctness of the vowel [ɛ], so the addition of this vowel into the disyllable adds to the expectation that kɪkɛ vs. keke would be easier than kɪ vs. ke. The results did not conform to this expectation.

Experiment 2, on the other hand, did provide some potential evidence for a perceptual advantage of harmony, showing that perception is significantly worse, and in some cases barely above chance, in most disharmonic contexts. There are three potential explanations for this result: (1) disharmony impedes the perception of ATR contrasts, (2) perception is more difficult for nonce words that are not possible words in the language, or (3) both vowels differ between A and B in the harmonic condition, but only one vowel does in the disharmonic conditions. This last option also connects to the potential of identity effects, since the harmonic disyllables in Experiment 2 had identical vowels in both syllables of each stimulus, while the disharmonic ones did not.

The second explanation is not plausible considering that disharmonic disyllabic forms are attested in compounds and sometimes in cliticization in Dàgáárè, as exemplified in Table 4. Since listeners are used to hearing such combinations in compounds, these nonce words would in fact be possible in the language, and so listeners should be able to adequately distinguish them. That said, we do not necessarily know how listeners are processing these disharmonic stimuli; it is possible that they have a processing cost due to not being possible as non-compound words in the language. As for the third possible interpretation, it still comes down to disharmony negatively affecting perception, as there are the same number of contrasts in the disharmonic conditions as in the monosyllabic one (i.e., one opportunity to hear [ɪ] versus [e]), yet performance is nonetheless worse in the disharmonic conditions. For example, when listeners hear [kɪke] versus [keke], they have the first syllable to hear the distinction between the two words, while the second syllable is the same in both words. This means that, in theory, they have the same amount of information to process the [ɪ] versus [e] distinction as in the forms [kɪ] versus [ke], yet they nonetheless perform worse on such forms. As such, the third possible explanation is effectively equivalent to the first, with the caveat that memory (disyllables being longer than monosyllables means more to hold in memory before a decision) and identity (harmonic disyllables in Experiment 2 had two identical syllables) may have an effect. Note that both of these points still point to a potential advantage of harmony and/or disadvantage of disharmony.

Thus, Experiment 2 suggests that disharmonic sequences, even though they exist in the language, make perception of difficult ATR/height contrasts even more difficult. This connects to the idea that harmony is due to listeners misperceiving disharmonic sequences. However, they seem to misperceive such sequences primarily in a way that is inconsistent with the phonological patterns. Specifically, listeners have the hardest time with a pair that differs in both ATR and height. If listeners have a particular tendency to misperceive [ɪ] as [e] and vice versa in harmonic contexts, then it is unclear why [ɪ] harmonizes to [i] rather than to [e]. This suggests that phonologization of misperception, as discussed by Ohala (1994), may not be the right explanation for the link between perception and ATR harmony. For example, if listeners are more likely to mishear [ɪ…e] as [e…e] than as [i…e], then it is the former that we would expect to see as the result of harmony, if harmony is the result of hypocorrection. Instead, the relationship between perception and harmony seems to be modulated through avoidance of misperception, similar to the suggestion for rounding harmony by Kaun (2004). If [i] and [ɛ] are the maximally distinct vowels, then their inclusion in a form should allow greater differentiation; a language allowing just [ɪ…ɛ] and [i…e] will have fewer misperception issues, since ATR values can be heard on [ɛ] and [i] respectively. While performance on differentiating these forms was not significantly better than for monosyllables in Experiment 1, that is expected, given that the finding for Experiment 2 was that harmony does not help perception but disharmony hurts. Nonetheless, this idea does have some issues. First, [ɪ…ɪ] and [e…e] are allowed in the language and are difficult to differentiate, as found particularly in Experiment 1. This fact makes any perceptual explanation for harmony quite difficult to motivate. Additionally, some languages have been said to have merged high RTR vowels with mid ATR vowels but either kept them distinct for the purposes of harmony (e.g., Aziza, 2008 on Urhobo; Omamor, 1988 on Okpe) or allophonically continue to harmonize them (e.g., Andersen, 1999 on Mayak). If the difficulty in differentiating that pair motivates harmony, then we would expect such a merger to make harmony redundant, and so it is odd that it would so consistently remain part of the grammar after such mergers. However, as ATR is already known to be different across languages, it may be that such languages have different contrasts that are perceptually difficult. More research is needed on perception in other ATR systems.

Given these considerations about [ɪ…ɪ] and [e…e], it is worth returning here to the issue of the difference between this pair in Experiment 1 versus Experiment 2, where the accuracy rates in distinguishing [ɪ…ɪ] versus [e…e] were 57% and 67% respectively. It is unclear what caused this difference, but it is worth exploring further in the future to understand why these sounds are allowed in ATR harmony languages, despite being difficult to distinguish. One possible explanation for the difference is that this was the hardest pair in Experiment 1, where all stimuli were harmonic, but not in Experiment 2, which included disharmonic stimuli. Perhaps the relative difficulty of words to differentiate had an effect on the results. Alternatively, it may be because many participants completed both experiments, so the experience with this pair that they had in Experiment 1 may have improved their ability to distinguish it in Experiment 2. Of course, in a real-world situation, semantic context may also help in differentiating real words with [ɪ…ɪ] and [e…e], so such considerations are also important in understanding the role of perception in the motivations for ATR harmony.

Returning to the issue of languages that are said to have merged high RTR vowels with mid ATR vowels (e.g., Aziza, 2008 on Urhobo; Omamor, 1988 on Okpe; Elugbe, 1983 on Edoid languages and proto-Edoid; Andersen, 1999 on Mayak), it is worth noting that, while our experiments are unable to provide any motivation for why these languages continue to show the ATR harmony systems they do after the merger, the experiments do potentially help us to understand why such mergers happen in the first place. As Casali (1995, p. 119) suggests, auditory similarity to and acoustic overlap with neighbouring vowels in the system could motivate the merging of high RTR vowels with mid ATR vowels. While the acoustic overlap is well-established and it is widely reported that [ɪ] and [e] are hard to distinguish, we noted in Section 1 that these perceptual comments are primarily anecdotal. Our study is among the first to confirm this fact experimentally with native speakers. The fact that it is common cross-linguistically for languages with ATR systems to merge these vowels could derive from the difficulty that even native speakers appear to have with their perceptual discrimination.

It is worth noting the standard orthography of most Ghanaian languages, including Dàgáárè, represents high RTR and mid ATR vowels the same way, as “e” and “o” for front and back respectively (cf. Bodomo, 1997). Thus, [ɪ] and [e] have the same spelling in Dàgáárè, which may affect native speaker perception and judgement. A plausible explanation for low accuracy in harmonic forms is the orthographic representation of the vowels [ɪ] and [e] with the letter “e”. By considering the phone-grapheme association of the letter “e”, the listeners might have perceived phonetically harmonic forms (e.g., [e]-[i]) as disharmonic (e.g., [ɪ]-[i]) and vice versa. However, participants were not shown any orthography, so it is unclear whether the spelling would have had any effect.

As mentioned in Section 2.2, there is very little existing baseline in the literature for understanding ATR perception and its possible role in ATR harmony systems. This study helps to provide a baseline about ATR perception, including which vowels are most difficult to distinguish for native speakers of a language with ATR, the difference between perception of ATR in monosyllables versus disyllables, and the effects of harmony and disharmony. While there remain many further questions to explore, as discussed in Section 5.5, this study makes an important contribution that can later be built on to answer some of the more nuanced questions of how perception might motivate ATR harmony. This is particularly true given the dearth of perception studies on ATR, and on African languages more generally, discussed in more detail in the next subsection.

5.4. Perception in African languages

Perception studies on African languages (on any property, not just ATR) are incredibly sparse. Based on a systematic literature review of literature since 2000, we are aware of a total of only eight published, peer-reviewed journal articles directly testing perception of African languages by native speakers in that time frame. At the word level, lexical tone perception has been investigated in Dinka (Remijsen, 2013), Yoruba (Harrison, 2000), Mambila (Connell, 2000), and Shona (Kadyamusuma, 2012), and there are also studies of tonal spreading in Bemba verb forms (Kula & Braun, 2015) and vowel length contrasts in Civili (Ndinga-Koumba-Binza & Roux, 2009). At the sentence level, perception of prosody and/or semantic focus has been explored in Akan (Genzel & Kügler, 2020) and Sepedi (Turco & Zerbian, 2021). In addition to the sound-related monolingual perception articles mentioned above, there have been a handful of studies on word perception by Twi (L1)-English (L2) bilinguals (e.g., Neumann & Nkrumah, 2019). There is also one very recently published perception study that has recently been conducted on African languages (Rose et al., 2023), and a small number of studies published more than two decades ago (e.g., Fulop et al., 1998 for Degema; La Velle, 1974, Harrison, 1996, and Hombert, 1976 for Yoruba; Bladon, Clark, & Mickey, 1987 for Shona). Of these studies, only Rose et al. (2023) and Fulop et al. (1998) are on ATR perception.

Given the dearth of research in perception on African languages, the present study represents a much-needed addition to the literature. Many more studies of this sort are necessary in order to further the representation of African languages in the literature and to thoroughly investigate the phonetic and phonological properties of these languages. In particular, our results are quite different from what Rose et al. (2023) found for ATR perception in Akan, emphasizing the need for more studies on more languages. Many further questions remain about the perception of ATR and the possible perceptual underpinnings of ATR harmony, which cannot be answered without further perceptual work on more languages.

5.5. Future directions