2.1 Data corpus

The majority of the acoustic data for this study comes from the Suomen Murrekirja (‘dialect book’) (Lyytikäinen et al. Reference Lyytikäinen, Rekunen and Yli-Paavola2013), an online repository of Finnish recordings. The Murrekirja is a digital corpus of about 500 recordings that are tagged with the city or town of origin. ^{Footnote 5} Each of these locations is also labeled with a dialect area. The recordings consist of one person talking for between two and five minutes, usually telling some kind of story—for example, about how life was when they were young, or about a bear hunter, or a wedding. The result is casual and dialectal speech, which is optimal for finding vowel insertion, a non-standard feature. The birth years of the speakers included in this study ranged from 1874 to 1905, with age at the time of recording ranging from 60 to 93. Gender was not indicated in the corpus, but was discernible from speaker names and the recordings (see Table 1 for more detailed information).

Data was collected from the locations listed in Table 1, where the database only includes one recording per location. The locations are marked in Figure 3, where Savo locations are marked with a circle and Pohjanmaa locations are marked with a diamond. Locations were selected from the list based on multiple criteria, including audio quality and presence of ${\rm{CV}}{{\rm{C}}_2}.{{\rm{C}}_3}{\rm{V}}$ words; I also made an effort to include locations from different regions of the dialect area, so as to not effectively be analyzing one small geographic region’s tendencies. Data from seven locations from each region was included, in total 14 recordings from the dialect corpus.

In addition to this corpus, a native speaker of the Oulu dialect provided me with a recording of them reading a text that they wrote to showcase Oulu inserted vowels (i.e., with the deliberate inclusion of words that show insertion). The text and recording were created without my input and before the conceptualization of this paper, and as such were not influenced by the hypotheses of this study. With this recording, the Pohjanmaa dialect has eight recordings, for a total of 15 recordings included in the dataset.

Two types of words were labeled for use in the quantitative analysis. The first, which I will refer to as triggering tokens, included words with triggering sequences in the correct prosodic position, such as lehmä ‘cow’ and kylmä ‘cold’. Triggering consonant sequences beyond the second mora were only included if they were in the second mora position of the second element of a compound, such as nykypolvelle [nyky+polvelle] ‘today’s generation.all’—i.e., a word like vadelma ‘raspberry’ would be excluded. In addition, /rC/ sequences were excluded from this category due to their controversial status, and instead will be discussed qualitatively. Although the exact consonant sequences could not be controlled for in the corpus, the samples are relatively balanced across dialect. Both dialects had similar proportions of different ${{\rm{C}}_2}{{\rm{C}}_3}$ sequences: /hC/ (Savo 18, Pohjanmaa 15); /lC/ (Savo 42, Pohjanmaa 36); and /nh/ (Savo 4, Pohjanmaa 7).

The second type of words (henceforth referred to as baseline tokens) were of the shape ${\rm{C}}{{\rm{V}}_1}{\rm{C}}{{\rm{V}}_2}{\rm{(X)}}$ . These were used to set a baseline of the expected qualities and durations of uncontroversially phonological vowels in V₁ and V₂ position. As V₁ and V₂ of a single word are directly compared, words with /i/ were only included if both V₁ and V₂ were /i/, such as in the word niminen ‘so-named’. This exclusion was made in order to avoid effects of intrinsic vowel duration muddying the data, as Wiik (Reference Wiik1965) noted that /i/ was significantly shorter than even other high vowels.

The full summary of the resulting corpus is presented in Table 2, with details on the number of tokens and lemmas in each location. Lemmas could be split based on trigger vs. baseline status; for example the word ulos ‘outside.lat’ was counted as a lemma for the baseline tokens, while a related form ulkona ‘outside.ess’ was counted as a separate lemma for the triggering tokens. Additional forms of ulko- were subsumed under the appropriate lemma—e.g. ulkopuolella ‘on the outside’ was counted as a member of the ulkona lemma. In all, the Savo dialect had 64 tokens of 29 triggering lemmas and 48 tokens of 42 baseline lemmas; the Pohjanmaa dialect had 60 tokens of 27 triggering lemmas and 65 tokens of 33 baseline lemmas. ^{Footnote 6} In total, the corpus had 124 tokens of 46 triggering lemmas, and 120 tokens of 67 baseline lemmas. Thus, while the corpus is small and may present some problems of generalizability, it does contain a fairly wide range of different words as well as some repetitions of those words to account for variability between productions and forms.

Table 2. Locations used in the acoustic corpus study, with the number of tokens (T) and lemmas (L) for the Triggering (TR) and Baseline (BL) words. Gender, year of birth (YOB), and age at time of recording are given for each speaker. The asterisk indicates the recording that was provided independently of the corpus

Table 3. A summary of the results of the diagnostic tests. Letters indicate if the characteristic is true of S(avo) dialects and/or P(ohjanmaa) dialects

2.2 Labeling and processing

As the Murrekirja does not include transcripts of the sound files, the help of native speakers of Finnish was enlisted to transcribe the recordings. The resulting transcripts were used to verify the presence of triggering and baseline tokens. Based on these transcripts, sound files were segmented using Praat (Boersma and Weenink Reference Boersma and Weenink2017); all triggering tokens were included unless sound quality issues prevented quantitative analysis (e.g., phrase-final devoicing of an entire word, low signal to noise ratio making segmentation impossible). Triggering tokens were judged to have a vowel or not with both auditory and visual (referencing both the waveform and the spectrogram) confirmation. If present, inserted vowels were marked in an interval separate from ${{\rm{C}}_2}$ ; if no vowel was judged to be present, only the consonant was marked. In order to not bias the marking of the length of the inserted vowel, segmentation of each sound file was done without knowing from which region the recording originated.

The resulting TextGrids were then processed in R (R Core Team 2019), using a script that gathered data on the identity and duration of each segment, using the package rPraat (Bořil and Skarnitzl Reference Bořil, Skarnitzl, Sojka, Horák, Kopeček and Pala2016). This data structure was then fed to a Praat script written by the author to obtain the vowel quality of V₁ and ${{\rm{V}}_i}/{{\rm{V}}_2}$ . Data analysis was carried out in R. Linear mixed effects models were implemented with the package lme4 (Bates et al. Reference Bates, Maechler, Bolker and Walker2014), with speaker as a random effect. These models were built incrementally and compared with likelihood ratio tests.

2.3 Measures and hypotheses

As in Section 1, in presenting the results of the acoustic study I will focus on the diagnostic criteria for inserted vowels provided by Hall (Reference Hall2006). Three of the diagnostics are quantifiable in this data:

1. 1. Frequency of insertion in triggering sequences. Is a vowel inserted consistently (indicating phonological status), or does its existence vary (indicating phonetic status)?

2. 2. Inserted vowel duration. For this analysis, I compare the duration ratio between ${{\rm{V}}_i}$ and V₁ (i.e., divide ${{\rm{V}}_i}$ by V₁) in triggering tokens to the ratio between V₂ and V₁ (i.e., divide V₂ by V₁) in baseline tokens. Are inserted vowels as long as underlying V₂ (indicating phonological status), or are they shorter and/or more variable (indicating phonetic status)?

3. 3. Inserted vowel quality. For this analysis, I again compare ${{\rm{V}}_i}$ to V₁ in triggering tokens, and V₂ to V₁ in baseline tokens that have matching vowel qualities. Is the Euclidean distance between ${{\rm{V}}_i}$ and V₁ on the one hand and V₂ and V₁ on the other comparable (indicating phonological status), or is the quality of ${{\rm{V}}_i}$ less predictable (indicating phonetic status)?

2.4 Consistency of insertion: mixed results

In this section, I examine the rates of insertion in triggering tokens, where any vocalic element is considered positive for insertion, regardless of duration. There is cross-dialect variation in the consistency of insertion, which indicates different degrees of phonologization in each dialect. Insertion is much more consistent in Pohjanmaa dialects (59 of 60 triggering tokens showed insertion) than in Savo dialects (47 out of 64 tokens showed insertion). This difference is statistically significant (Pearson’s ${\chi ^2} = $ 12.14, p = 0.0002). Insertion rates for each dialect are illustrated in Figure 4.

Figure 3. Map of Finnish dialects used in this study, adapted from Lyytikäinen et al. (Reference Lyytikäinen, Rekunen and Yli-Paavola2013) and cut to focus on relevant dialect regions (Pohjanmaa dialects in light gray region; Savo dialects in dark gray region). Location tags correspond to the letters in Table 1.

Figure 4. Summary of triggering C₂C₃ sequences in the data, and the frequency of vowel insertion in each. C₂C₃ sequences are collapsed into three categories: /hC/, /lC/, and /nC/ (all /nh/).

In the Pohjanmaa dialect, the single token that failed to insert was a token of the word vanha ‘old’, produced by the speaker from Lestijärvi. Although Kettunen (Reference Kettunen1940) specifically separates the word vanha (and other /nh/ sequences) from the main vowel insertion map, I included it in the “triggering” types of tokens because the area indicated on the map was roughly coextensive with the more general types of vowel insertion, and the /nh/ sequence follows the phonological requirements for insertion. In the corpus data, there are an additional four tokens of vanha (or some derived form); all exhibited insertion. As the Lestijärvi speaker did not produce this word or consonant sequence again, it is impossible to tell if this is an issue of variability within one speaker, or if this sequence simply never triggers insertion for them.

In contrast, almost all Savo locations had at least one triggering sequence fail to insert. The major locus of variability is in the /hC/ sequences: out of 18 /hC/ tokens, 13 (72.2%) did not show insertion. Five of these tokens were produced by the Varpaisjärvi speaker, who appears to never insert in /hC/ sequences—out of five words with an /hC/ sequence, none showed insertion. However, it does not seem to be the case that /hC/ sequences just fail to trigger insertion in some Savo dialects, as the speaker from Kajaani sometimes inserted in /hC/ sequences and sometimes did not. This variability was present even within the same word, illustrated in Figure 5. In this figure are three instances of the word kahvi ‘coffee’: the first (Figure 5a), has a vocalic portion with strong formant structure between the /h/ and /v/; the second (Figure 5b) has a vocalic portion between the /h/ and /v/ with less clear formant structure; and finally, the third (Figure 5c) has no such vocalic element, instead going straight from the /h/ to the /v/.

Figure 5. Three examples of kahvi ‘coffee’ by Kajaani (Savo) speaker. Inserted vowels marked with Vi.

There was also some variability in /lC/ sequences, though all tokens that failed to insert were from the speaker from Ranua. In this case, there was some variability in /lk/ sequences; some tokens of the same words showed insertion, while others did not. This is illustrated in Figure 6. In this figure, there are two instances of the word ulkona ‘outside.ess’, accompanied by another word with the same consonant sequence, jalkaa ‘leg.part’ (Figure 6a). In the first instance of ulkona (Figure 6b), there is a clear—if relatively short—vocalic element after the /l/; in the second (Figure 6c), there is no inserted vowel. The example of jalkaa (labeled “jalakaa”) provides a good comparison for the previous two tokens: the inserted vowel is clearly present and quite long in comparison to V₁. Thus, although there is some variability in /lC/ clusters as well, it is limited to one speaker and clearly a case of token-to-token variability, rather than overall failure to trigger.

Figure 6. Three /lk/ sequences by a Ranua (Savo) speaker. (a) jalakaa (long inserted vowel); (b) ul(u)kona (relatively short insertion); (c) ulkona (no insertion).

Of all ${{\rm{C}}_2}{{\rm{C}}_3}$ sequences, those involving /h/ would be the most likely to lag behind in phonologization. As described in Section 1, FVI only occurs in ${{\rm{C}}_2}{{\rm{C}}_3}$ sequences where ${{\rm{C}}_2}$ has phonetic voicing. In sequences like /hm/ and /hv/, the /h/ frequently receives some voicing by being surrounded by sonorants, which provides an environment for FVI. However, a voiced [ɦ] is very vowel-like; Ladefoged and Maddieson (Reference Ladefoged and Maddieson1996) noted that it has been described as a breathy-voiced version of a vowel, and Keating (Reference Keating1988) observed that the vocal tract during a glottal fricative simply takes the shape of whatever segments are around it. Thus, even if an excrescent vowel occurs after the voiced [ɦ], the articulatory and acoustic similarity between the [ɦ] and vowel would make it more difficult to perceive the vowel as a separate segment. This, in turn, would impede phonologization: unlike in, for example, /lC/ sequences, where insertion would have been irregular but perception consistent, /hC/ sequences would have shown variability both in insertion and in perception.

In sum, the two dialects show variation in the consistency of insertion, with an interaction between dialect and ${{\rm{C}}_2}$ . Pohjanmaa shows nearly 100% insertion, which is a trait of phonological epenthesis. Savo insertion is more variable, indicating phonetic excrescence, but the variability is centered on the /hC/ sequences. This suggests that the two dialects are at different stages of phonologization: Pohjanmaa has fully phonologized the inserted vowel, while Savo is lagging behind, specifically in the /hC/ sequences.

2.5 Duration of inserted vowels indicates phonological status

In this section, I compare the duration of inserted vowels in triggering tokens to the duration of the underlying V₂ in baseline tokens. Since inserted vowels are typically copies of V₁, it was possible to normalize duration across speakers, speech rate, and phrase position by dividing the duration of V₂ or ${{\rm{V}}_i}$ by the duration of V₁ (in all cases an underlying vowel), producing the ratio of V₂ (or ${{\rm{V}}_i}$ ) to V₁ in a particular token. In total, there were 65 baseline words for the Pohjanmaa dialect, and 55 baseline words for the Savo dialect. This analysis only includes triggering tokens where insertion was observed; thus, the triggering tokens where insertion was predicted but did not occur are excluded.

Overall, the duration of inserted vowels patterns the same as underlying V₂, despite small statistically significant differences. When including all tokens with insertion, the addition of word type significantly improves the model ( ${\chi ^2}(1) = 6.17,{\rm{p}} = 0.01$ ), where the ratio between inserted ${{\rm{V}}_i}$ and V₁ ( $\beta $ = 1.44, SE = 0.06) is significantly smaller than the ratio between underlying V₂ and V₁ ( $\beta $ = 1.61, SE = 0.07). The difference in ratio between word types is quite small, which does not suggest a difference in phonological status between inserted and underlying vowels. In addition, the ratios for both word types is approximately 1.5, as illustrated in Figure 7. This is in fact expected in these dialects, which exhibit so-called “second-mora lengthening” (Suomi and Ylitalo Reference Suomi and Ylitalo2004). In these dialects, V₂ of a ${\rm{C}}{{\rm{V}}_1}{\rm{C}}{{\rm{V}}_2}{\rm{(X)}}$ word is approximately 1.5 times as long as V₁ (further detail on second-mora lengthening will be provided in Section 3). The presence of this lengthening points again to phonological status: not only is the inserted vowel as long as an underlying V₂—it is that long because it is treated as a second mora and lengthened accordingly.

The major exception to this appears to be the /hC/ tokens that do show insertion in Savo dialects, and these tokens drive the difference in ratio between ${{\rm{V}}_2}/{{\rm{V}}_1}$ and ${{\rm{V}}_i}/{{\rm{V}}_1}$ . In general, the vowels in Savo /hC/ sequences are quite short, even compared to vowels in Pohjanmaa /hC/ sequences. This is illustrated in Figure 7, where almost all of the tokens in the lower tail of the Savo inserted vowel distrubtion are /hC/ tokens, marked with an “H”: of the five /hC/ tokens that did trigger insertion, four are at the very bottom of the distribution. When /hC/ tokens from Savo speakers are removed from the dataset, the addition of word type as a factor does not significantly improve the model ( ${\chi ^2}(1) = 3.53,{\rm{p}} = 0.06$ ). The addition of dialect as a factor does not significantly improve either model ( ${\chi ^2}(1) = 0.0001,{\rm{p}} = 0.99$ with all tokens included; ${\chi ^2}(1) = 0.14,{\rm{p}} = 0.71$ with Savo /hC/ tokens removed), nor does the interaction between dialect and word type ( ${\chi ^2}(1) = 2.16,{\rm{p}} = 0.14$ with all tokens included; ${\chi ^2}(1) = 0.51,{\rm{p}} = 0.78$ with Savo /hC/ tokens removed). Thus, for Pohjanmaa dialects, both consistency of insertion and the duration of the inserted vowel indicate that FVI is a phonological phenomenon. For Savo dialects, the evidence is mixed: /lC/ sequences very consistently produce long vowels, while /hC/ sequences inconsistently show insertion, and the inserted vowels that exist are generally much shorter than underlying vowels.

2.6 Vowel quality: copied vowels in both dialects

In this section, I compare the Euclidean distance (in Hz) between V₁ and inserted V₂ (in triggering tokens) and V₁ and underlying V₂ (in baseline tokens). Formant measures were taken from the middle 20 ms of each vowel using Praat’s Get Formant functions. Euclidean distances were calculated using the F1 and F2 of each vowel. In order to compare directly between inserting words (where ${{\rm{V}}_i}$ is a copy of V₁) and baseline words, only baseline tokens with the same vowel quality for both V₁ and V₂ were included in these models. As there are very few tokens per person, and not all peripheral vowel qualities could be extracted for every speaker, it was impossible to do a study-wide normalization; however, as for the duration measures, the comparison of vowels of the same quality within one word effectively normalizes the vowel space across the study.

In both dialects, inserted vowels are no further from V₁ than an underlying V₂ of the same quality. The addition of word type does not significantly improve the fit of the model ( ${\chi ^2}(1) = 0.007,{\rm{p}} = 0.93$ ), and the inclusion of dialect as a second fixed effect also does not significantly improve the model ( ${\chi ^2}(1) = 0.50,{\rm{p}} = 0.48$ ). The interaction between dialect and word type also does not significantly improve the model ( ${\chi ^2}(1) = 0.04,{\rm{p}} = 0.83$ ). This indicates that in both dialects, inserted vowels are typically copies of V₁, and not either schwa or some intermediate vowel. This is as reported by Kettunen (Reference Kettunen1940)—the only region with intermediate quality vowels is the southernmost region of Savonia, and no speakers in this sample are from that particular area.

As Euclidean distances convey only absolute value and not direction, it is also prudent to consider the peripherality of inserted vs. underlying vowels. Although a statistical analysis is not possible, the plots in Figure 8 illustrate that inserted vowels are not, on the whole, more centralized than the underlying V₁. If anything, Savo inserted vowels appear to be largely more peripheralized, and Pohjanmaa vowels appear to be shifted slightly back. These two facts combined indicate that the inserted vowels have undergone a high degree of phonologization.

Figure 7. A violin plot showing the ratio of V₂ to V₁ durations, divided by dialect. For triggering tokens (CVC₂C₃VX), V₂ is the inserted vowel; for baseline tokens (CVCVX), V₂ is an underlying V₂. A horizontal line is at 1.5, the typical ratio from second-mora lengthening. The location of /hC/ tokens is marked with H.

Figure 8. Vowel spaces for triggering tokens in Pohjanmaa and Savo dialects, represented as convex polygons that encompass all tokens in the respective datasets. The polygons of V _i have no outline and the polygons of V₁ have a solid outline. Individual vowel values are labeled with their quality; A stands in place of ä (/æ/).

The one exception to inserted vowels being a copy of V₁ is the vowel inserted in /lj/ sequences. As described in Section 1, inserted vowels are reported to be [i] when ${{\rm{C}}_3}$ is /j/, rather than a copy of V₁. In the dataset overall, there are 7 tokens with insertion where ${{\rm{C}}_3}$ is /j/; the mean Euclidean distance between ${{\rm{V}}_i}$ and V₁ is 507.09 Hz when ${{\rm{C}}_3}$ is /j/, compared to a mean of 147.99 Hz for the other 119 tokens. The movement is overall towards the /i/ quality, as pictured in Figure 9. As Hall (Reference Hall2006) argues that only excrescent vowels are affected by surrounding consonants, this last pattern of [i] in /Cj/ sequences indicates, minimally, an excrescent origin, if not enduring phonetic status.

Figure 9. Vowel space showing the average movement from V₁ to V _i when C₃ is /j/ (both dialects combined).

2.7 Qualitative analysis

In this section, I present a qualitative analysis of the insertion phenomenon. First, I discuss some tokens of “blocking” sequences that appear to have some sort of vocalic interval, which indicates excrescence; this is followed by a brief discussion of the participation of the inserted vowels in phonological processes.

2.7.1 Evidence from blocking sequences

In this section, I present some evidence from “blocking” tokens (i.e., /rC/ and voiceless ${{\rm{C}}_2}$ sequences) that contrast with the apparently phonologized inserted vowels, and thus provide support for an excrescent origin. Both blocking types present their own impediments to phonologization. First, heterorganic voiceless sequences could potentially have an excrescent vowel, but due to the lack of a voicing gesture in the consonants, the gap would be voiceless. This is in fact similar to Dep(f), the constraint proposed by Harrikari (Reference Harrikari1999) to address the same gap; for Harrikari, the insertion of the feature [+voice] was prohibited in [-voice][-voice] consonant sequences, and thus a vowel could only be inserted if there is at least one voiced consonant. This voicelessness would discourage reinterpretation as a vowel, particularly in a language (such as Finnish) that does not have word-internal devoiced vowels. However, while the excrescent vowel may not be voiced, there still appears to be a fairly long release in at least some sequences of voiceless consonants, as illustrated for the word pitkiä in Figure 10. In this word, there is a 35 ms space of aspiration-like noise after the release burst for the /t/, reminiscent of the aspiration described for Sierra Popoluca (Gafos Reference Gafos2002).

Figure 10. The word pitkiä ‘long-part.pl’ with a 35 ms aspirated interval of C₂, by a Kuivaniemi (Pohjanmaa) speaker.

Yet another barrier to phonologization could be the attributability of the vocalic interval to ${{\rm{C}}_2}$ , for example when ${{\rm{C}}_2}$ is an /r/, which is typically accompanied by short vocalic intervals even without additional gestural underlap. As noted in Section 1, there is a debate in the literature over the insertion status of /rC/ sequences, though the proponents of each side both believe in a fully phonological origin of the inserted vowels. For this paper, I will take the stance that vowel insertion in heterorganic /rC/ sequences is present, but the vowel is still fully excrescent and thus inconsistent, and resists phonologization due to factors I will discuss below.

The first piece of evidence comes from consultation with native speakers of inserting dialects. Although Suomi (Reference Suomi2000) reported that his Oulu consultants categorically denied insertion, the six native inserting speakers that I consulted on this question have no such consensus. One respondent felt that it was not possible to insert a vowel in any /rC/ sequences at all; another thought that insertion in /rC/ sequences was perhaps a bit strange, but not impossible. Interestingly, three respondents thought that insertion was possible only in /rj/ sequences (i.e., not /rk/ or /rv/ sequences or other sequences with similar places of articulation in ${{\rm{C}}_3}$ ), but that the vowel is “weak” (example given: kirja ‘book’ > [kiria], without the [j]); a fourth thought that there would probably be no insertion in /rk/ or /rv/ sequences except possibly in fast speech (example given: surkeaa ‘pitiful.part’ > [sur(u)kiaa] ^{Footnote 7} ), but that there definitely was an inserted [i] in /rj/ sequences (examples given: korjata ‘to fix’ > [korijata], kirja ‘book’ > [kirija]). This disagreement among speakers suggests that there may sometimes be vowel-like intervals in /rC/ sequences, but they are certainly not deliberately produced.

Accordingly, there are some /rC/ tokens in the corpus that appear to have a longer-than-expected open interval between the closed portion of /r/ and the next consonant. As previously described, /r/ in coda position is canonically realized as a trill (Suomi et al. Reference Suomi, Toivanen and Ylitalo2008). In some tokens in the corpus, the last open portion of the trill cycle lasts long enough such that it may be considered vowel-like. An example of this is provided in Figure 11a. In this token, a Kajaani (Savo) speaker produces the word kerjätä ‘to beg’ with a full trill (three closures), and the last open portion before the glide is 40 ms long, which is short compared to V₁ (82 ms) but long compared to the other open portions in the trill (both less than 20 ms long).

Figure 11. Three examples of /rC/ sequences in the corpus. (a) The word kerjätä ‘to beg’ with an excrescent vocalic portion following a trilled /r/, by a Kajaani (Savo) speaker; (b) the word arvasi ‘guess.3sg.past’, by a Lestijärvi (Pohjanmaa) speaker; (c) the word varmaan ‘surely’, by a Maaninka (Savo) speaker.

It should be noted that the example provided in Figure 11a is an /rj/ sequence, which is the sequence that four of the six native speakers thought was likely to have a vowel. This apparent “exception” may be the result of a conflict between the articulatory postures necessary for /r/ and /j/ that makes gestural underlap more likely: the tongue body has to be braced in a certain position so that the tongue tip is free to vibrate, and this posture is incompatible with the high front constriction necessary for /j/. Thus, there is almost necessarily an interval of time where the tongue posture is no longer producing a trill, but also has not fully achieved the /j/ constriction; this interval could be interpreted either as part of the /r/ or as a separate vowel segment. In contrast, there would be much less conflict between the postures for /r/ and /k/, for example, and no conflict at all between /r/ and any labial consonant. Another example of a trill, this time with a bilabial ${{\rm{C}}_3}$ and no long vocalic portion, is illustrated in Figure 11b. In this token, a Lestijärvi (Pohjanmaa) speaker produces the word arvasi ‘guess.3sg.past’ with a two-cycle trill; in contrast with the token in Figure 11a, the final open portion is approximately equal in duration to the first open portion.

There are, however, some sequences other than /rj/ that seem to have vowel-like intervals between the consonants, particularly when the /r/ is produced as a tap, rather than a trill. It is unclear if the tokens of /rC/ words with taps are idiosyncratic to the speaker or to a particular production of a word, but in these types of /rC/ tokens especially there seems to be a tendency to have a vocalic interval following. Taps are very short and previous work has described a vocalic portion on either side of the tap, as the body of the tongue must brace for the tap motion (Gibson et al. Reference Gibson, Sotiropoulou, Tobin and Gafos2017); any delay of the following consonant constriction would simply extend this unconstricted interval, creating an excrescent vowel. One example of this type vocalic interval is provided in Figure 11c (from Maaninka, Savo dialect). Here, the word varmaan ‘surely’ has a 47 ms vocalic portion after the single closure for /r/. This vocalic portion is comparable in duration to some of the excrescent /hC/ tokens in Savo dialects, but is much shorter than would be expected from a vowel in second mora position.

Overall, /rC/ sequences exhibit inconsistent insertion, and tokens that do have insertion uniformly have short vocalic intervals. The variability of insertion seems to have impeded the phonologization of the vowel, similarly to /hC/ sequences in the Savo dialect. Furthermore, as /r/ inherently comes with short open intervals (whether tapped or trilled), phonologization may be further inhibited by the possibility of attributing these excrescent vowels to the consonant itself, which is not available for consonants like /l/ and /n/. These characteristics address the debate in the literature: vowels sometimes appear (as Harrikari Reference Harrikari1999 argues), and sometimes do not (as Suomi Reference Suomi1990 argues); as they are not phonological and appear as the result of gestural underlap, they are less likely to be perceived by speakers. A failure to be perceived by speakers does not necessarily mean that the vowels do not occur, but it does indicate that they are not phonologized.

2.7.2 Participation in phonological processes

Finally, Finnish inserted vowels have an interesting dual patterning in other phonological and morphophonological processes. In her 1999 analysis, Harrikari treated FVI as a phonological repair, indicating phonological status. However, she also noted that the inserted vowel appears to be invisible to allomorphy that is sensitive to the number of syllables. In Finnish, there is variability in the plural cases, where for example the trisyllabic word lakana ‘sheet’ is realized as [lakanoita] in the partitive plural (Anttila Reference Anttila, Chambers and Schilling2002, Reference Anttila2010). Disyllables, on the other hand, are realized without the /t/: the word kala ‘fish’ is realized as [kaloja] in the plural partitive and *[kaloita] is ungrammatical. Words that are disyllables without insertion and trisyllables with insertion function as disyllables in this respect: e.g. the plural partitive of ahma ‘wolverine’ ([ahama] in the inserted form) can only be realized as [ahamoja], not *[ahamoita]. The impossibility of the trisyllabic forms has been confirmed by the panel of native inserting speakers consulted for this paper—and unlike the possibility of inserting vowels in /rC/ forms, they are all in agreement on this fact.

On the other hand, although the inserted vowels are never stressed, they do appear to count towards footing. In Finnish, primary stress is fixed on the first syllable, and secondary stress falls on every other syllable thereafter modulo an attraction to weight (Kiparsky Reference Kiparsky, Mannine and Nelson2003; Karvonen Reference Karvonen2005). Final syllables that follow an unstressed syllable are also unstressed unless they are heavy, in which case they optionally receive secondary stress. However, when comparing the footing rules described by Karvonen (Reference Karvonen2005) to Pohjanmaa judgments (provided by a subset of the panel of native speakers that have an intuition for secondary stress), it is clear that the inserted vowel is visible to footing:

The relevant forms here are (8c) and (8d), which both have four syllables in the inserted form and have two trochees. There are two things to notice in this data: first, that the consulted speakers explicitly included the syllable boundaries that indicated that the inserted vowel can form a nucleus, and second, that these inserted segments count for the footing process. Thus, at least for these speakers, the vowels are phonologically active in stress assignment. The apparent contradictory patterning of the inserted vowel between morphophonological processes (where the vowel is not treated as a nucleus) and stress assignment (where the vowel is treated as a nucleus) supports an excrescent origin with phonologization: the plural partitive patterns for words with vowel insertion were fossilized prior to the phonologization of vowel insertion, but post-lexical stress processes treat the vowel as a true phonological unit.

3.1 SML and the dialectal distribution of FVI

Second-mora lengthening (SML) is precisely as the name implies: in dialects that have SML, the segment in second mora position is lengthened. Hints of SML are visible in other Finnic languages in the so-called “half-long vowel”, which describes the extended duration ^{Footnote 8} of V₂ in ${\rm{C}}{{\rm{V}}_1}{\rm{(C)}}{{\rm{V}}_2}{\rm{(X)}}$ words (see Prince Reference Prince1980; Asu and Teras Reference Asu and Teras2009 for Estonian (est), Kiparsky Reference Kiparsky, Kehrein, Köhnlein, Boersma and van Oostendorp2016 for Livonian (liv), and Gordon Reference Gordon2009 for Ingrian (izh)). This half-long vowel has also been documented as a feature of central and northern Finnish dialects since at least the late 19th century:

More recently, a number of phonetic studies (Suomi and Ylitalo Reference Suomi and Ylitalo2004; Suomi Reference Suomi2009; Ylitalo Reference Ylitalo2009) have shown that the half-long vowel is still robust in central and northern dialects, and is preserved by university-educated female speakers of Northern Finnish even in laboratory-elicited speech (Ylitalo Reference Ylitalo2004). Moreover, these studies have also shown that the half-long vowel phenomenon has been generalized to the second-mora position, thus generating the name “second-mora lengthening” (Suomi and Ylitalo Reference Suomi and Ylitalo2004; Suomi Reference Suomi2009; Ylitalo Reference Ylitalo2009). This phenomenon affects consonants in second mora position as well as vowels—that is, precisely the consonant that triggers vowel insertion. The effects of SML are summarized and exemplified below:

Linking FVI explicitly to SML neatly accounts for the dialectal distribution of FVI: crucially, the dialects of Finnish that do not exhibit SML also do not have FVI (Kettunen Reference Kettunen1940; Wiik and Lehiste Reference Wiik and Lehiste1968; Spahr Reference Spahr2012). If FVI were viewed as an entirely separate innovation, it would be one that developed independently across the major East-West Finnish dialect split, as Savo dialects are in the Eastern group and Pohjanmaa dialects are in the Western group. In addition, at least some of the Western Finnish dialects spoken in the Lapland region (the “Far North” dialects) also exhibit FVI (Kettunen Reference Kettunen1940) as well as SML (Spahr Reference Spahr2012); however, due to the lack of experimental and corpus data available from these regions, they have been excluded from the present paper.

If we instead view SML as a necessary precursor to FVI, the major innovation is limited to the hämäläismurteet, which are the subset of Western dialects spoken by over two million people in the major regions of the south of Finland. SML is not currently documented in these dialects (see e.g. Ylitalo Reference Ylitalo2009 for the dialect spoken in Tampere, as well as Wiik and Lehiste Reference Wiik and Lehiste1968 for a more general survey), and thus must have been lost minimally before the phonologization of the inserted vowel occurred. An investigation of why some dialects with SML develop FVI but not others is beyond the scope of this paper; however, in some cases SML may be produced by a different timing structure that would not generate FVI. For example, the dialect region of Turku has been documented to have the half-long vowel (Wiik and Lehiste Reference Wiik and Lehiste1968; Ylitalo Reference Ylitalo2009) but not vowel insertion. However, Ylitalo (Reference Ylitalo2009) argues that SML in Turku dialects is not the same as SML in the Oulu (Pohjanmaa) dialect: while the Oulu dialect achieves SML through lengthening the second mora, the Turku dialect achieves a SML through shortening the first mora. Thus, the lengthening process would not be targeting ${{\rm{C}}_2}$ , and accordingly would not generate vowel insertion.

3.2 SML and the prosodic distribution of FVI

As Hall’s (Reference Hall2006) model of excrescent vowels is rooted in articulatory gestures in the Articulatory Phonology tradition, a gestural model of SML and FVI is necessary to reap the full benefits—to create gestural underlap, one must first have a gestural model. A fully formalized and computationally implemented model is beyond the scope of this paper; however, I discuss here the components of a gestural model that I believe would generate both SML and FVI. A link between SML and FVI addresses the positional limitation, which has previously not been satisfactorily accounted for. The relevant environment is ${\rm{CV}}{{\rm{C}}_2}.{{\rm{C}}_3}{\rm{V(X)}}$ , and not later ${{\rm{C}}_2}{{\rm{C}}_3}$ sequences—this is precisely when ${{\rm{C}}_2}$ would be targeted by SML. This makes a slightly more complicated argument for gestural underlap. Typically, excrescence via gestural underlap occurs whenever that particular sequence of consonants occurs, or is limited by position within a syllable (e.g., excrescence occurs only in onsets, or only in codas, or heterosyllabically); in these dialects of Finnish, underlap is conditioned by second mora status.

In order to account for the very particular environments, I propose the existence of an oscillator in Finnish that corresponds to a bimoraic foot, which is coupled both to segmental gestures as well as to a boundary $\pi $ gesture. This configuration is illustrated in Figure 12. In Articulatory Phonology, an oscillator is simply “any process that tends to repeat itself regularly” (O’Dell and Nieminen Reference O’Dell and Nieminen2009: 2), after harmonic oscillators in physics. Articulatory gestures are typically analyzed as oscillators in Articulatory Phonology, as in repetitive speech (such as [mamama]) there would be an oscillatory motion of the articulators. In addition, oscillators for larger prosodic units have also been proposed (O’Dell and Nieminen Reference O’Dell and Nieminen2009), such as for the mora, the syllable, and the foot.

Figure 12. Schematics for the bimoraic oscillator (black trajectory) coupled to CVCV(X) and CVC(X) gestural scores. Note that in both cases the coda /l/ is stretched due to its position in the coda, but the $\pi $ gesture warps the bimoraic oscillator further than the stretched coda can go (figures left-aligned for ease of comparison).

The “bimoraic foot” descriptor of this proposal is simply a gestural interpretation of Suomi’s (Reference Suomi2009) bimoraic “locus of accentual lengthening”. In his study, Suomi found that segments in the first two moras of a word were lengthened relative to comparable segments in later positions; the second mora, in particular, was greatly affected, producing SML. Karlin (Reference Karlin2015) also argued for a bimoraic foot at the beginning of a word in Central and Northern dialects of Finnish, citing both SML as an instantiation as domain-final lengthening, as well as another dialectal repair that effectively prevents a bimoraic foot from dividing a syllable (i.e., CVCV: > CVC:V:).

The production of SML itself relies on an additional $\pi $ gesture (Byrd and Saltzman Reference Byrd and Saltzman2003), at the end of the bimoraic oscillator. As the alternative name for these gestures, “clock-slowing gesture”, implies, a $\pi $ gesture at prosodic boundaries slows the progress of some oscillator, effectively locally lengthening the gestures. These gestures have been invoked to describe processes such as phrase-final lengthening and the warping of gestural timing relations near prosodic boundaries (Byrd and Saltzman Reference Byrd and Saltzman2003; Katsika et al. Reference Katsika, Krivokapić, Mooshammer, Tiede and Goldstein2014). In dialects with SML, there appears to be an active prosodic boundary at the end of the bimoraic foot, which comes with boundary-associated lengthening. A clock-slowing gesture here would locally slow the progress of the bimoraic oscillator, producing SML.

An oscillator may also be coupled to another oscillator, either of the same type (e.g., segment-to-segment coupling) or different (e.g., foot-to-segment). In this model, the bimoraic oscillator is hierarchically coupled to the gestures that make up the segmental content. An oscillator has a particular preferred frequency, but its actual timing may be influenced by being coupled to another oscillator. For example, O’Dell and Nieminen (Reference O’Dell and Nieminen1999) proposed that a driving force of so-called “stress-timed” and “syllable-timed” languages is in fact the dominance of one type of oscillator over another: in stress-timed languages, an oscillator corresponding to stress groups dominates over a syllable oscillator, while in syllable-timed languages, the dominance relationship is reversed.

Dominance may additionally be affected by other properties of oscillators, such as mass or stiffness. The remaining question for this proposal is why consonant sequences show underlap and resultant excrescence, while vowels simply stretch to accommodate this newly-created “half-long” spot. I attribute this failure to stretch to high gestural stiffness, which effectively describes an oscillator’s resistance to deformation. In Articulatory Phonology, stiffness has been proposed as one of the defining parameters of gestures (Browman and Goldstein Reference Browman and Goldstein1989, Reference Browman and Goldstein1990); gestures with high stiffness have a shorter cycle and higher amplitude (such as consonants with high degrees of closure), while gestures with low stiffness have a longer cycle and lower displacement amplitude (such as vowels; see e.g. Fuchs et al. Reference Fuchs, Perrier and Hartinger2011 for kinematic data and mathematical models). In this particular situation, the bimoraic oscillator is coupled with sufficient dominance to the segmental gestures to force vowel gestures to deform, as they have low resistance to displacement, but not consonant gestures, as they are highly resistant to deformation. Thus, the bimoraic foot oscillator is deformed by the $\pi $ gesture, but cannot sufficiently deform the consonant gesture, effectively producing a gap where the bimoraic oscillator is still active without full articulatory closure.

Finally, although for this paper I have focused on ${\rm{CV}}{{\rm{C}}_2}.{{\rm{C}}_3}{\rm{V(X)}}$ forms, it should be noted that insertion is also present when ${{\rm{C}}_3}$ is a geminate, as in the word helppo ‘easy’ > [helep.po]. The gestural proposal still works for these words—the $\pi $ gesture is still targeting the second mora, but in this case, the second mora simply has two associated closure gestures. The same effects thus apply: the closure gesture for ${{\rm{C}}_2}$ attempts to stretch but fails, producing a gap between that gesture and the following ${{\rm{C}}_3}$ ; the lengthening of the second gesture may also fail but the homorganic closure that follows closes the gap. This type of “non-local” lengthening has been documented in phrase-final lengthening, where, for example, lengthening effects are found in penultimate syllables, rather than just final segment or even syllable (Turk and Shattuck-Hufnagel Reference Turk and Shattuck-Hufnagel2007); as the $\pi $ gesture has been proposed as the major driver of phrase-final lengthening, the current application fits with existing data.

3.3 The phonologization of phonetic detail

Finally, there is still work to be done regarding the process of phonologization of the excrescent vowels. Hall (Reference Hall2006) cites several cases of fully phonologized excrescent vowels, including related cases such as Lule Saami (Engstrand Reference Engstrand1987) and Lapua (southern Pohjanmaa) Finnish (Harms Reference Harms1976). Ohala (Reference Ohala and Jones1993) has argued for language change due to processes of hypocorrection, where phonetic effects are reinterpreted as intentional, as in tonogenesis where the raised F0 that accompanies voiceless consonants is reinterpreted as the primary cue. In this case, phonologization essentially entails a process of turning nothing into something—i.e., a gap with no active constriction is reanalyzed as a full segment with an intentional gesture. Hall (Reference Hall2006: 424) also notes that phonologization is neither “an inevitable fate for intrusive vowels, nor does it happen automatically upon their reaching some threshold of phonetic duration”. In this section, I turn to the process of phonologization of FVI in Central and Northern dialects, and discuss possible influences on phonologization as derived from the coupled oscillator model, and as evident from the acoustic corpus.

The two end stages of this process are illustrated in the schemata in Figure 13, using a gestural score that correlates to the previously described coupled oscillator model. It is worth noting that this model does not necessarily predict no lengthening of the consonants, simply insufficient lengthening to form sufficient constriction for the entire “half-long” time warping of the bimoraic oscillator (a slight lengthening of /l/ is depicted in Figure 13a). More specifically, the model predicts that stiffer consonants would be most likely to create gestural gaps, as they would be the most resistant to deformation; less stiff gestures (such as the fricative /h/) would be less resistant and thus more able to follow the stretching of the foot-level oscillator. This aligns with the /lC/ and /nC/ exhibiting full phonologization in both dialects, as they both involve full closure at the tongue tip and are thus very stiff. It also aligns with the /hC/ sequences lagging behind in the Savo dialect, as /h/ is more open, less stiff, and thus more prone to deformation.

Figure 13. Schemata of a simplified gestural score corresponding to (a) an excrescent vowel and (b) a fully phonologized epenthetic vowel in the word kolme.

These stiffness differences would have a twofold effect on the phonologization process. First, one factor that one would expect to influence phonologization is the consistency with which insertion occurs. If /h/ is sufficiently flexible to only occasionally create gaps, then those gaps would be less likely to be interpreted as deliberate. Similarly, the behavior of /rC/ sequences also suggests that /r/ is less stiff than /l/ and /n/; this may be literally true in the stiffness of the tongue tissue (which has also been suggested to be linked to gestural stiffness, as in Fuchs et al. Reference Fuchs, Perrier and Hartinger2011) despite the full closures. However, the /r/ case is even more complex, as it is influenced by articulatory and aerodynamic constraints, as well as inherent open intervals, none of which affect either /l/ or /n/. Second, although Hall (Reference Hall2006) noted that duration by itself is not sufficient to predict phonological status (as evidenced by Scots Gaelic), longer gaps are more likely to be interpreted as intentional. As /l/ and /n/ are less easily deformed, the small amount that they do deform would leave a bigger gap than the more flexible /h/. Thus, this model predicts some cross-segment variability and thus does not rule out different rates of phonologization. Although the corpus data overall reflects a fairly late stage of phonologization, the Savo dialect in particular hints at stiffness-correlated differences in phonologization. A future study may look for dialects with true excrescence across the board and examine the duration of the inserted vowel with different ${{\rm{C}}_2}$ .