2.1. Meters

For present purposes, the meters of the Rigveda are classified according to the number of syllables per line (in the restored text), being eight, eleven, or twelve.Footnote ¹² Other line lengths, which are very infrequent, are set aside. Moreover, only the regular versions of these three line types are considered, each of which has a particular line structure. Less frequent line types are excluded, despite having the requisite number of syllables.Footnote ¹³ Further, the Vālakhilya section (0.8% of the corpus) is excluded, being an interpolation. Finally, repeated lines are excluded after their first instances (reducing the corpus by another 6.7%). After these exclusions, 33,229 lines remain (11,234 eight-syllable, 15,620 eleven-syllable, and 6,375 twelve-syllable), comprising the regular dimeter and trimeter corpus employed here.Footnote ¹⁴ Dimeter refers to the eight-syllable line, which is divided evenly into opening and cadence, albeit with no caesura. Trimeter refers to the eleven- and twelve-syllable lines, which feature a break between the opening and cadence and a caesura after the fourth or fifth position (both positions are available in both meters; Arnold Reference Arnold1905). For transparency, I refer to the three line types as meters 8, 11, and 12.

2.2. Strong and weak positions

Some tests below invoke positional strength, operationalized as follows. As is evident from Figure 1, in all three meters, weight-mapping is the most regular in the cadence. Regularity manifests both in strictness—each position being nearly categorically heavy or light—and also in the rhythmic, binary alternation between lights and heavies. Based on these criteria, the cadence comprises the final four positions of meters 8 and 11 and the final five of meter 12 (Lubotsky Reference Lubotsky, Semeka and Pankratov1995:515). Adopting terminology from generative metrics (e.g. Halle & Keyser Reference Halle and Keyser1971, Kiparsky Reference Kiparsky1977, Prince Reference Prince, Kiparsky and Youmans1989, Hanson & Kiparsky Reference Hanson and Kiparsky1996, Hayes et al. Reference Hayes, Wilson and Shisko2012), I refer to cadential positions that are normally filled by heavies as Strong (S) and those that are normally filled by lights as Weak (W). The remaining, pre-cadential positions are coded as Other (O). The three meters then have the forms in 1.

Two caveats are due about this treatment of strong vs. weak positions. First, in all meters, line-final position is conventionally understood to be indifferent to weight. As Figure 1 reveals, it is heavy roughly eighty percent of the time across meters. This applies even to meter 11, where I just referred to final position as weak. In generative metrics, positional strength is abstract—one might refer to an iambic or trochaic cadence—and final indifference is a license. More importantly, cross-linguistically, final indifference applies only to weight, not to stress or tone (Ryan Reference Ryan2019b:139). Thus, if accent were regulated by Vedic meter, a priori, one would not expect line-final position to be exempt. A second caveat is that the restriction of strong/weak to the cadence is not meant to suggest that there are no weight tendencies or strong/weak positions elsewhere in the line. But such tendencies are weaker (see Figure 1) and more complex: for example, in meters 11 and 12, caesura is usually immediately followed by a pair of lights, but the location of caesura is not fixed. Another reason to focus on the cadence is that if accent were metrically relevant, one might expect its effects to be clearest there, given that the cadence is metrically the strictest part of the line not just in Vedic, but cross-linguistically (ibid., deCastro-Arrazola Reference deCastro-Arrazola2018).

An anonymous reviewer raises the valid concern that even though weight is more strongly regulated in the cadence, hypothetically, that might not be true for accent, if accent is regulated. I therefore rerun every test with an alternative scheme in which S and W positions are coded throughout the line, as described in Section 3.6. For cohesion, I postpone presenting those tests until Section 3.6 rather than interleaving them throughout the article, but they yield qualitatively the same (negative) results as the original tests as coded in example 1.

2.3. Part of speech

At various points, I control for part of speech, based on annotations by Hellwig et al. (Reference Hellwig, Hettrich, Modi and Pinkal2018) (see also Hellwig et al. Reference Hellwig, Scarlata, Ackermann and Widmer2020). These top-level labels include noun, adjective, verb, adverb, verbal noun, and (in fewer than one percent of cases) other. Different parts of speech are distributed differently in the line (e.g. later in the line, verbs become more frequent relative to nouns) and also have different inherent accentual tendencies (e.g. verbs are more likely to be finally accented than nouns). Additionally, based on the same source, I distinguish between vocative and non-vocative nouns, as vocatives have special accentual behavior: their accent (if any) can only be word-initial. Vocatives are usually unaccented, but when line- or clause-initial (or when chained with other vocatives that are such), they receive a default initial accent. This is the only morphosyntactic category for which the location of accent in the word is determined by position in the line. Vocatives are therefore often excluded below.

2.4. Word shape

Some tests control for each word’s metrifiable shape, which affects where the word can appear in the line (Devine & Stephens Reference Devine and Stephens1994:74–79, Gunkel Reference Gunkel2010, Ryan Reference Ryan2011:419). To a first approximation, a word’s shape can be regarded as its syllable-weight template (e.g. light-heavy-light). However, more detail is needed for the peripheral syllables, whose weights are context-sensitive. Here, word shape encodes (i) whether all non-final syllables are heavy or light, (ii) whether the word begins with a vowel or consonant, and (iii) the weight class of the ultima (four levels). Vowel-initiality matters because a word-initial vowel may coalesce with a preceding vowel, in which case the resulting syllable is heavy.Footnote ¹⁵ For the ultima, I define four weight classes based on the rime, namely, VVC, V, V:, and VC/e/o. VVC (where VV is any long vowel or diphthong) is always heavy. V is usually light, but may glide or coalesce with a following vowel, potentially forming a heavy. V: is any long monophthong (not counting e and o, which, while long, function as diphthongs); before a consonant, a V: ultima is heavy; before a vowel, it coalesces, glides, or stands in hiatus (in which case it usually shortens). Finally, VC/e/o (where V is short) is heavy before a consonant and light before a vowel. In this manner, every word is coded for its metrifiable shape, which constrains its distribution in meter (e.g. agním ‘Agni’ is V-heavy-VC/e/o, mahā́n ‘great’ is C-light-VVC, and sukr˳tyáyā ‘by good deeds’ is C-light-heavy-light-V:).

2.5. Word type

Finally, word shape and part of speech can be combined into a single index termed word type. For example, agním ‘Agni’ is type V-heavy-VC/e/o-noun. These combinations are able to capture any idiosyncratic effects of word shape in different categories, both in terms of the distribution of words within the line and in terms of accentual tendencies within words.

3.1. Disyllable distribution

A disyllable may be initially or finally accented, or unaccented. The present test considers only accented disyllables, so as not to bias the results by the distribution of word-level accentedness (i.e. whether the word contains an accent or not), which, as discussed, declines across the line for reasons independent of meter. Furthermore, I exclude vocatives for the reasons discussed (Section 2). If accent were regulated by meter, one might expect to see the ratio of initially- to finally-accented disyllables vary systematically across positions. In particular, if accent behaved directionally like weight, one might expect it to be preferred in strong positions and/or avoided in weak positions.Footnote ¹⁶ That is, one might expect the ratio of σ́σ to σσ́ to be higher spanning strong-weak as opposed to weak-strong positions (cf. Hayes et al. Reference Hayes, Wilson and Shisko2012:701 on stress in English).

Figure 2 shows the percentage of accented disyllables that are initially accented in each position of the three meters. Position here refers to the locus of the word’s initial syllable. Strong and weak positions (as defined in Section 2) are shaded and striped, respectively. Error bars are 95% confidence intervals using Wilson (Reference Wilson1927) scores. While the plots for meters 8 and 12 might suggest a weak tendency for accent-meter alignment within the cadence, the cadential rates are generally in the range of the rates elsewhere in the line. Moreover, meter 11, the most frequent meter, shows no such pattern. An additional caveat is that the plots do not control for weight. For instance, in general, heavy-light words are more likely to be initially accented than light-heavy words (52% vs. 43%, including indeterminate ultimas in both groups). This alone might inflate the incidence of accent in strong positions relative to weak positions. Therefore, it is important to control for weight.

Figure 2.

Percentage of disyllables starting in each position that are initially as opposed to finally accented. Solid-shaded positions are strong, striped positions weak.

Mixed-effects logistic regression is used to assess the overall effect of positional strength on accent while controlling for weight (word shape) and part of speech. The model predicts whether each accented disyllable $ \left(n=42,540\right) $ is initially accented. The lone fixed-effect predictor is position type (strong, weak, or other). As random effects, I employ the specific metrical position (e.g. ‘11-5’), each with its own intercept (i.e. bias for accent, modulo other factors). I also include random intercepts for word type, which combines word shape with part of speech (Section 2). Finally, the model includes random slopes for position type conditioned on word type, which corrects for any idiosyncratic tendencies of particular word types to be initially accented in particular position types. The regression formula is given in 2.Footnote ¹⁷

The results (fit by glmmTMB; Brooks et al. Reference Brooks, Kristensen, van Benthem, Magnusson, Berg, Nielsen, Skaug, Mächler and Bolker2017) are provided in Table 2. The $ p $ -values are not significant (i.e. all are $ p>0.05 $ ), but the model does not directly test strong vs. weak; rather, it compares strong and weak separately to the baseline (pre-cadence positions). To check all three possible contrasts, a post hoc Tukey’s test is conducted using emmeans (Lenth Reference Lenth2024). This test confirms that strong and weak positions are neither distinct from the baseline nor from each other (lowest $ p=0.67 $ ).

Table 2.

Regression table for fixed effects in the disyllable model.

In conclusion, with weight being controlled, there is no significant tendency for the accents of disyllables to be aligned with strong positions or avoided in weak ones.

3.2. Monosyllable distribution

A second, independent test is conducted, now on monosyllables. Since the location of accent cannot vary in a monosyllable, the present test differs from the previous one in not being confined to accented words. Rather, the distribution of accented monosyllables is compared to that of unaccented monosyllables, in either case excluding vocatives. (The remaining unaccented monosyllables comprise clitics and occasionally verbs.) The overall rates are shown in Figure 3. Because a line-initial word must be accented, the rate is 100% for that position. The general uptrend in accentedness from the second position to the middle of the line plausibly reflects the distribution of clitics, which frequently follow the line-initial word (Hale Reference Hale and Watkins1987). Moreover, the spikes at positions five and six of meters 11 and 12 reflect the avoidance of an enclitic immediately following caesura.

Figure 3.

Percentage of monosyllables in each position that are accented.

At any rate, the interest here is the cadence, and whether accent correlates with strong vs. weak positions. The figure suggests no such correlation for meters 8 and 12, though some cycling is observed in meter 11. Nevertheless, the figure does not control for weight or part of speech. Under a weight- and category-controlled regression, as in Section 3.1 (except now excluding line-initial position, which is categorically accented here), no pairing of position types (strong, weak, or pre-cadential) is significant (lowest Tukey’s test $ p=0.83 $ ; $ n=\mathrm{18,045} $ ).

That said, Figure 3 reveals another way in which accent might interact with meter: across meters, the accent rate drops in line-final position, regardless of whether that position is strong or weak. Indeed, when a fixed effect for line finality is added to the previous model, it is significant $ \left(p=0.002\right) $ , while position type remains non-significant (lowest $ p=0.36 $ ). The effect of line-finality will be confirmed in Section 3.4.

3.3. Other word shapes: an omnibus model

Monosyllables (Section 3.2) and disyllables (Section 3.1) together account for thirty-eight percent of the syllables in the corpus. The majority of syllables—and hence potential data—are furnished by longer words, to which this section turns. While the previous tests had dichotomous outcomes (initial vs. final accent for disyllables; accented vs. not for monosyllables), for longer words, more outcomes are possible, making a dichotomous test inappropriate. For example, in a trisyllable, accent (if any) could be initial, medial, or final.

One approach to handling longer words would be to switch from a by-word model (as in previous tests) to a by-syllable model, in which the outcome is the accentuation of each syllable in the corpus, restoring dichotomy. Alas, such an approach should not be used to judge significance, as it violates an assumption of logistic regression (Sonderegger Reference Sonderegger2023): within each word, the accentuation of each syllable is not an independent datum, since words typically have at most a single accent.

I therefore employ a word-based Monte Carlo method, using random permutations to obtain $ p $ -values.Footnote ¹⁸ Let a (single-accented) word of any length be denoted strong-aligned if its accent coincides with a strong position, and likewise for other position types, as illustrated for some possible trisyllable alignments to meter 8 in Table 3.

Table 3.

Some possible alignments of a trisyllable’s accent with meter 8.

For a given word shape, one can compute its corpus-wide strong-alignment rate, which is 21.2% for single-accented, non-vocative trisyllables in the regular meter corpus. The question is whether this observed rate is higher or lower than expected by chance. A chance baseline is computed by randomly permuting (shuffling) the accent patterns of trisyllables while leaving their loci intact. Under one such permutation, 20.9% of trisyllables are strong-aligned. The near-identity of the real and false (randomized) rates suggests that the poets were not sensitive to accent when localizing trisyllables. However, the method just outlined is incomplete, since the false rate varies on each computation; any single estimate might be an outlier. This is where the Monte Carlo method comes in. The false rate is recomputed a large number of times (here, 1,000). The $ p $ -value is then the proportion of trials in which the false rate exceeds the real rate, which is 0.026 in this case. While this $ p $ -value is ostensibly significant, this simple illustration did not control for weight or part of speech. Moreover, the effect size is tiny (real 21.2% vs. mean false rate of 21.0%).

The permutation method generalizes to any number of word shapes of any specificity, provided that accent patterns are exchanged only between shapemates. To illustrate, Table 4 shows three arbitrary word types (specified as in Section 2) with their real and false (one trial) strong-alignment rates. For each trial, the weighted sum of real rates (here, 26.6%) is compared to that of false rates (here, 26.2%). The Monte Carlo $ p $ -value is the proportion of trials in which the overall false rate exceeds the overall real rate (here, $ p=0.61 $ ). Once again, this example is only an illustration of the method, not a full result.

Table 4.

A sample of word types and their real vs. false alignments on one trial.

Full results for all single-accent, non-vocative words of three or more syllables in the regular meter corpus $ \left(n=\mathrm{49,692}\right) $ are given in Table 5. Five hypothetical position types are tested. The first row pits strong positions (as defined above) against all other positions, finding that the real rate of accent alignment is virtually identical to the expected (mean false) rate. Weak positions are likewise non-significant.Footnote ¹⁹

Table 5.

Final Monte Carlo results for various position types.

Line-finality, however, is significant in Table 5, in that over all 1,000 trials, the real rate is less than the expected rate $ \left(p<0.001\right) $ .Footnote ²⁰ Because strong and weak positions overlap with line-final position, they were retested in the final two rows of Table 5 by counting only non-final strong/weak positions. With this adjustment, weak remains non-significant, but strong becomes significant, albeit with a small effect size (19.9% vs. 19.4%).

The same method can be used to gauge not only position types, but specific metrical positions, which is useful for seeing whether the effects just discussed are consistent across positions of each type, serving as cross-validation. In Figure 4, each dot is the odds ratio of observed to expected alignment (where the expected rate is the mean of 1,000 permutations), and 95% confidence intervals are the [2.5, 97.5] percentiles of trials. I note the following. First, the overall range of odds ratios is small (most being between 0.9 and 1.1, where one represents no difference). Second, most of the error bars overlap with one (and even more would with any penalty for multiple tests). Third, and most importantly, other than final position, position types (strong or weak) do not behave consistently as classes. In particular, 43% of strong positions are above the baseline (60% if line-final strong positions are excluded). Similarly, 50% of weak positions are above the baseline (60% if the line-final weak position is excluded).

Figure 4.

Monte Carlo observed vs. expected rates for each position as odds ratios.

To conclude, this section, treating polysyllables, finds a significant and reliable effect for line-final position such that across meters, final accent is avoided relative to the shape- and category-controlled baseline. As for strong positions, although this section finds that they tend significantly to align with accent in the aggregate, the tendency is small (half a percent) and inconsistent across individual strong positions. I therefore take the present results to agree with those of Sections 3.1–3.2, which likewise find no reliable effect of strong or weak positions, but support a line-final effect.

3.4. The avoidance of line-final accent

As just discussed, across meters, line-final position significantly disfavors accent, even when controlling for the expected rates of specific contexts using random effects. This section follows up on two items. First, it reinforces that the effect is largely consistent across word types, and therefore unlikely to be an epiphenomenon of specific words or grammatical types (which happen to be oxytone) being underrepresented line-finally for independent reasons. Second, I test whether Vedic prose also exhibits the preferential avoidance of line-final accent, in which case it would not be a specifically poetic phenomenon.

In the metrical corpus, for single-accented, non-vocative words of at least two syllables, Figure 5 shows the ten most frequent word types (as defined in Section 2). Each type is accompanied by its rate of word-final accent in two contexts, namely, non-line-final (left, lighter bar) and line-final (right, darker bar). For the first eight types, the line-final rate is lower, while the final two rows are essentially ties. The model in 3 confirms the generalization for all word types (including those not shown in Figure 5). Word-final accent significantly reduces the likelihood of a word’s appearing line-finally (odds ratio 0.78, $ p<0.0001 $ , $ n=\mathrm{92,232} $ ). The random intercepts for word type (e.g. C-heavy-VC/e/o-noun) as well as for specific word identity (e.g. agním) reinforce that this effect is not being driven by particular items.Footnote ²¹

Figure 5.

Ten most frequent word types, each with its rate of word-final accent when non-line-final (light) vs. line-final (dark). In almost every case, the line-final rate is lower.

I turn now to Vedic prose in order to test whether line-final accent is also disfavored in non-metrical contexts. As an accented Vedic prose corpus, I employ the prose sections of the Taittirīya Saṃhitā (also known as the Black Yajurveda) and Taittirīya Brāhmaṇa.Footnote ²² To (largely) filter out metered material, I exclude all lines of eight, eleven, or twelve syllables. After further removing duplicate lines, 20,850 lines remain. The analysis here considers only accented words of at least two syllables, and further excludes line-initial words.Footnote ²³

As Figure 6 reveals, tonal N on F inality (see Section 5) is once again in strong effect. Because weight is unregulated in prose, word shape is replaced by word size in syllables. For all word sizes, word-final accent is significantly less likely line-finally than line-medially. Logistic regression (same formula as in example 3, except using word size in lieu of word type) confirms that the effect is general, and not being driven by specific words or word sizes (odds ratio 0.49, $ p<0.0001 $ , $ n=\mathrm{64,404} $ ). I conclude that while tonal Non Finality is robust in the Rigveda, it is not a metrical phenomenon, but more general the language.

Figure 6.

Word sizes (2–6 syllables) in Vedic prose, each with its rate of word-final accent when non-line-final (light) vs. line-final (dark). In every case, the line-final rate is lower.

A reader asks whether NonF inality also holds of word-level phonology, and it does, but only as a tendency in longer words. For disyllables, accent is roughly equally likely to be initial or final, but for increasingly long words, final accent is increasingly underrepresented (Sandell Reference Sandell2023a:425–28). For example, 14% of quadrisyllables (by type) have final accent (ibid.). To be sure, this word-level NonFinality does not directly explain the line-level NonFinality found in this section. The regression models include random effects for word type and identity, which means that regardless of how infrequent final accent is in words, it is even less frequent line-finally. For example, for a quadrisyllabic noun with a particular weight pattern, the model predicts a baseline rate of line-finality via the random intercepts. If one adds the information that the quadrisyllable is finally accented, that rate is adjusted downwards by the fixed effect.

Thus, the following type of explanation would not be compatible with this section’s results: line-final accent is underrepresented because longer words tend to be line-final, and they are unlikely to be finally accented. The random effects undermine such an explanation. Likewise, one cannot say that the effect is driven by unaccented words being more likely line-finally, since the model only uses accented words as data. A successful explanation would need to apply generally across word types (shapes and parts of speech). For example, perhaps there is a low boundary tone L% which is felt to be in tension with H] $ {}_{\iota } $ . I do not claim that this is the actual explanation, just that it is the type of explanation that is needed.

The reader also raises the point that Vedic poetry and prose are both high-register genres, in which sentences are carefully crafted, and might therefore be more responsive to euphony than ordinary language. It is true: I have shown that line- or sentence-level NonFinality is not confined to meter, but the effect might not hold for all registers.

3.5. The post-accentual fall (svarita) is also not a target of alignment

Throughout the article, I test meter’s alignment with the phonological locus of high-tone accent, termed udātta ‘raised’ and indicated in romanization by the acute. In phonetic realization, the udātta is rising, with the peak achieved either late in the syllable or on the following syllable, depending on the text/tradition (Macdonell Reference Macdonell1910:77, Beguš Reference Beguš, Jamison, Melchert and Vine2016:4). The post-udātta syllable (if not itself udātta or immediately pre-udātta) is realized with a sharp drop known as the svarita (ibid.), which Pāṇini 1.2.31 defines as samāhāra ‘compound’, in the sense of high + low. One could imagine that this fall is perceptually salient.

I will test whether svarita is better aligned with the meter than udātta presently, but find it a priori unlikely, for two reasons. First, svarita is merely an ‘allotone’, dependent on the phonological accent (udātta).Footnote ²⁴ It is common for accented words to lack the svarita altogether due to their context; svarita, unlike udātta, is a post-lexical, phrasal phenomenon.Footnote ²⁵ Second, svarita was almost certainly not the pitch peak at composition. Some recitational traditions still reach the peak on the udātta, including the padapāṭha version of the Rigveda recorded in Kerala in 1961 (Witzel Reference Witzel, Oettinger, Schaffner and Steer2020). Sandell (Reference Sandell2023a:294) concludes that ‘[g]eographically separate regions thus preserved the original Vedic intonational system without peak delay’.

Nevertheless, to be thorough, I test whether the svarita is better aligned to the meter than the udātta. Here, a word-shape-specific test is inappropriate, given svarita’s phrasal nature. I therefore consider overall alignment with strong vs. weak positions in the regular meter corpus, excluding line-final position. For udātta, the strong-to-weak ratio is 1.17 (15,571 to 13,289). For svarita, it is 0.94 (11,967 to 12,725). Svarita is not better aligned than the udātta.

3.6. An alternative, more inclusive coding scheme for positional strength

While weight is regulated more strictly in the cadence than elsewhere in the line, if accent were regulated, the same generalization might not hold for accent. Therefore, my original coding scheme for positional strength (Section 2), which recognizes strong and weak positions only within cadences (pre-cadence positions being ‘other’), might be overconservative. In this section, I rerun the tests above (disyllable, monosyllable, and omnibus), but now coding strength throughout (most of) the line, as in 4. The overall verdict remains negative: there is no discernible correlation between accent and positional strength.

For all three meters, the opening (i.e. first four positions) is, if anything, abstractly iambic (WSWS), as suggested by the distribution of weight in Figure 1. (I have confirmed that when word shape is controlled, the weight biases remain.) Meters 11 and 12 further exhibit a ‘break’ of three positions (5–7) between the opening and cadence, which I leave coded as ‘other’. While weight in Figure 1 might suggest that all three positions are weak, because there is no regular rhythmic alternation, and because caesura interrupts this span (in one of two possible positions), analyzing strength in this part of the line is less straightforward.

Under this new scheme, the disyllable test (Section 3.1) is non-significant ( $ \beta =-0.019,p=0.88 $ for strong vs. weak), as is the monosyllable test (Section 3.2) ( $ \beta =0.245,p=0.57 $ ). While monosyllable accentedness appears to alternate in openings in Figure 3, a general effect of position type is not supported (given also the random effects for word type). An effect might be supported if openings were analyzed separately from the rest of the line, but aside from concerns about fishing, such a test would not be probative: recall that the monosyllable test (unlike the other tests) considers only whether words are accented; as such, it largely reduces to a test of the distribution of clitichood. Clitics are richly constrained at the left periphery (e.g. Wackernagel’s law). Further, the trend in Figure 3 contradicts the posited iambic pattern. Finally, it would be surprising, perhaps universal-breaking, for the meter to regulate a prosodic feature in the opening but not the cadence.

The omnibus Monte Carlo tests for longer words (Section 3.3) are borderline here. For strong positions (coded as in example 4, but excluding line-final position as in Section 3.3), the alignment of the real corpus is 35.7%, vs. a mean of 35.4% over 1,000 permutations $ \left(p=0.018\right) $ . For (non-final) weak positions, the real rate is 34.0% vs. a mean of 33.7% for permutations $ \left(p=0.993\right) $ . In both cases, the difference is miniscule (0.3%). Moreover, for weak positions, the directionality is the opposite of the prediction: accent occurs in weak positions more often in the real corpus than it does in the permutations.

4.1. Accentual responsion

Thus far, I have focused on ways in which accent might interact with Vedic meter. Beyond meter, other poetic uses of accent are conceivable. This section considers whether accentual contours tend to agree (in whole or in part) between adjacent lines, a (possible) phenomenon that I refer to as accentual responsion.Footnote ²⁶ (‘Accentual rhyme’ would be another way of putting it.) Lubotsky (Reference Lubotsky, Semeka and Pankratov1995) proposes that the Rigveda exhibits various kinds of accentual correspondence between lines, exemplifying them but not testing them statistically.

I begin by extracting all couplets from the regular meter corpus (Section 2), where a couplet is operationalized as adjacent a and b lines, or c and d lines, provided that their length (and hence meter) is the same (see Gunkel & Ryan Reference Gunkel, Ryan, Gunkel and Hackstein2018 on the Rigvedic couplet). The resulting corpus comprises 13,400 couplets, eighty-one percent of the original corpus.Footnote ²⁷ Among these couplets, 0.98% agree (i.e. have repeated accent patterns). As a baseline of expectation, the couplets can be shuffled, such that each couplet-initial (a or c) line is paired with a randomly chosen line of the same length from another verse. As in Section 3.3, the shuffle can be iterated many times to increase the reliability of the estimate and obtain a $ p $ -value – the Monte Carlo method. In this case, for 1,000 trials, the mean false rate is 0.72% $ \left(p=0.001\right) $ .

While this $ p $ -value is ostensibly significant, it is flawed. The method fails to control for other kinds of parallelism that incidentally inflate accentual responsion. For example, consider 5.38.5 (lines c–d) in example 5 (translation Jamison & Brereton Reference Jamison and Brereton2014), one of the rare couplets with full agreement in accent. Whenever one or more words are repeated in the same position across lines—for any reason—it increases the likelihood that the lines as wholes will match in accent.Footnote ²⁸ Additionally, the couplet in 5 illustrates a less trivial point: even absent literal repetition, structural parallelism can induce accentual agreement. Here, both lines begin with a vocative noun. Because a line-initial vocative must be initially accented (Section 2), this grammatical parallelism entails accentual parallelism. The same would hold if the vocatives were line-final (both unaccented), or if both lines ended with principal-clause finite verbs (both unaccented), and so forth. Thus, even purely syntactic parallelism (which abounds; Gonda Reference Gonda1959, Klein Reference Klein2001, Reference Klein2002) stacks the deck in favor of accentual agreement relative to the randomized baseline.

It would be impossible to control for every conceivable kind of parallelism, but spurious parallelism can be reduced by excluding any couplet in which (i) any word is repeated between lines or (ii) the two lines have the same number of words. For consistency, I apply these criteria to both real and scrambled couplets. If the poets strive for accentual responsion per se, it should obtain independently of other kinds of parallelism. The criteria just stated only partially break parallelism; much of it will remain in the real couplets (e.g. syntactic parallelism between lines with different numbers of words, the same lexeme being repeated but with different endings, etc.). At any rate, even this partial reduction in parallelism is enough to eliminate the effect: among real couplets, the rate of responsion is now 0.58%; among scrambled couplets, 0.65% $ \left(p=0.87\right) $ .

So far, I have tested only for exact correspondence between lines. It is also conceivable that the poets favored partial correspondence (between parts of lines and/or skipping positions). To test this looser hypothesis, I adopt various strictness thresholds based on the Levenshtein edit distance between the two accent patterns (i.e. the number of substitutions needed to bring the patterns into alignment). Results for distances up to two are given in Table 6. (For greater distances, the baseline exceeds 50%.) Although one of the distances is marginally significant, the effect size is miniscule (0.36% higher than chance). Given further that it is sandwiched between non-significant results,Footnote ²⁹ on top of which uncaptured parallelism surely remains in the real couplets, I take the overall verdict to be negative.

Table 6.

Monte Carlo results for accentual responsion.

A simpler but more limited test focuses on specific word shapes in specific positions. For example, consider all couplets in which both lines begin with a disyllable $ \left(n=\mathrm{3,503}\right) $ . From these, I exclude any couplet in which the two disyllables are the same word, leaving 2,751 couplets. The question is whether the accent pattern of word A (from the first line) is predictive of that of word B (from the second). Based on the contingencies in Table 7, it is not (Fisher’s exact test odds ratio 0.91, $ p=0.21 $ ). In the table, ‘10’ refers to initial accent, ‘01’ to final accent. The conclusion is the same for line-final disyllables, not shown (odds ratio 0.87, $ p=0.31 $ ).

Table 7.

Accent patterns of successive line-initial disyllables.

Other kinds of accentual correspondence are logically possible. One could imagine, for instance, that responsion obtains between couplets rather than lines, in which case a would correspond with c, and b with d. I reran the test in Table 6 with this treatment (using a/c and b/d as couplets) and got similar results: while the $ p $ -values are borderline, the caveats to Table 6 apply.Footnote ³⁰ In conclusion, this section has tested some simple possible scenarios for accentual responsion, finding no support for the phenomenon in the Rigveda.

4.2. Accentual clash and lapse

Some languages avoid adjacent prominent syllables, a configuration known as clash (e.g.Tilsen Reference Tilsen, Fuchs, Weirich, Pape and Perrier2012:122f). In a tonal context, the Obligatory Contour Principle (OCP) penalizes adjacent identical tones (e.g. Hyman Reference Hyman, Goldsmith, Riggle and Alan2011b:494ff). Additionally, some languages avoid adjacent unstressed syllables, a configuration known as stress lapse (e.g. Gordon Reference Gordon2004). (Lapse is less likely to apply to tone, but I test it anyway.) Together, constraints against clash and lapse favor alternation in prominence. That is, if both *Clash and *Lapse are undominated, stress occurs on every other syllable. Because Vedic tonal accent is culminative (at most one per word, barring unusual cases), clash and lapse are moot within words. Nevertheless, it remains possible that they might emerge as preferences across words (reminiscent of the rhythm rule of English; e.g. Tilsen Reference Tilsen, Fuchs, Weirich, Pape and Perrier2012). Specifically, if clash/OCP were active, it might manifest in the underrepresentation of bigrams of the form σ*σ́ σ́* (i.e. adjacent accents across a word boundary). Likewise, if lapse were active, one might observe a dispreference for boundaries flanked by unaccented syllables.

Table 8 is set up like Table 7, except that now words A and B are adjacent words within the line. There is no contingency between the accent patterns of the two words (Fisher’s exact test odds ratio 1.01, $ p=0.76 $ ). Trisyllables (of type 100 vs. 001) were also tested, with no effect (odds ratio 0.87, $ p=0.13 $ ). In conclusion, I find no evidence for the OCP or accentual lapse in word arrangements in the Rigveda.

Table 8.

Accent patterns of successive disyllables.

4.3. Accentual formulas

Like clash and lapse, the question of accentual formulas concerns syntagmatic tendencies in accent arrangement, except now more generally: When arranging words, are there particular sequences of accent that the poets favor above chance? This section finds no such tendencies.

As an initial test, let a line’s accentual cadence be the accent pattern of its final five syllables.Footnote ³¹ In the regular meter corpus, all $ {2}^5=32 $ possible combinations are attested except 11111 (all-accented). The frequency of each such pattern in the real corpus (even when zero) is compared to that of ‘scrambled’ corpora. Each scrambled corpus $ \left(n=10\right) $ randomly permutes the word accent patterns of the real corpus while keeping the loci of word types (Section 2) fixed, such that the weight, boundary, and part of speech patterns are held constant. For each accentual cadence, an observed vs. expected (O/E) value is computed and significance-tested (exact binomial test, Bonferroni-adjusted for thirty-two tests). For example, pattern 10110 occurs 229 times in the real corpus and 2,211 times in the ten scrambles, making for an O/E of 1.04 (adjusted $ p=1 $ ). Of all thiery-two accentual cadences, only one has a positive, significant O/E, namely, 00110 (odds ratio 1.14, $ p=0.02 $ ). However, the effect size is small, the $ p $ -value is marginal, and the greater likelihood might simply reflect the avoidance of line-final accent that was already established (Section 3.4), which inflates the O/Es of all 0-final accentual cadences.

Since an accentual formula might hypothetically occur elsewhere in the line, I use the same method to test five-position sequences starting from line-initial position. Of thirty-two patterns (all of which are attested in the real corpus), none has an O/E of significantly greater than one. In conclusion, I find no evidence of accentual formulas in the Rigveda.

4.4. Accent-modulated weight mapping

This section tests for one final possible role of accent in Vedic meter, which is that weight is regulated more strictly in accented syllables. Typologically, this phenomenon has been observed for stress and weight. Ryan (Reference Ryan2017) refers to it as stress-modulated weight mapping. The meter of the Finnish Kalevala, for instance, is a quantitative trochaic tetrameter, such that odd positions are heavy and even positions light, especially later in the line. However, this mapping rule is stringently enforced only for stressed syllables; unstressed syllables are largely free to be heavy or light in any position (ibid., Kiparsky Reference Kiparsky and Gribble1968). One could imagine a similar tendency in Vedic, such that exceptions to the quantitative template might be less frequent under accent. From the regular meter corpus, I take only non-line-final, cadential strong and weak positions (Section 2). In these positions, the overall rate of exceptionality (i.e. light in strong or heavy in weak) is 3.3%. Considering only accented syllables in these positions, the rate is 4.2%; for unaccented syllables, it is 2.9%.Footnote ³² In sum, no regular effect of accentual modulation is supported.