Comparative Analysis of Majority Language Influence on North Sámi Prosody Using WaveNet-Based modeling

5.1 The extended DigiSámi corpus

To analyze the prosodic differences between the North Sámi areal varieties and mapping the majority language influence, we used the DigiSámi read speech corpus (Jokinen & Wilcock, 2014). It was collected from five locations traditionally inhabited by the Sámi: Anár, Avvil, and Ohcejohka in Northern Finland, and Guovdageaidnu and Kárášjohká in Northern Norway (see Figure 1). The DigiSami corpus is available via the CSC website (www.csc.fi) by contacting the author of the original data (Jokinen, 2018).

The task of the speakers was to read aloud Wikipedia articles about the Sámi languages and traditions. The corpus consists of speech data from 25 native (16 females) North Sámi speakers, with ages ranging from 16 to 65. The data were recorded using a 4-channel portable recorder (Roland Edirol R-4 Pro) and a condenser mini-lavalier microphones (AKG C 417L). The speech data also originally contain sentence level annotations—for a more detailed description of the data collection and corpus, see Jokinen, 2014, 2018; Jokinen et al., 2017). According to the questionnaire information described by Jokinen (2018), most speakers used North Sámi as the main language of communication when interacting with other people and there were no remarkable differences in the language skills between the participants.

To extend the original DigiSámi speech corpus, an additional corpus (the Extended DigiSámi corpus) with nearly the same texts read by different speakers was collected by the authors in 2018 in Guovdageaidnu and Oulu, following the ethical guidelines of the University of Helsinki. There were altogether 148 individual sentences in the text material of the Extended DigiSámi corpus. Speech data from seven (four females) more speakers was added to the Extended DigiSámi corpus (age range 22–64). The five speakers recorded in Guovdageaidnu were all from Guovdageaidnu except one, originally from Kárášjohká. The speakers recorded in Oulu were originally from Ohcejohka and Avvil. To make the two corpora as compatible as possible, the recording process followed similar protocol as in the Extended DigiSámi corpus (Jokinen & Wilcock, 2014), with the inclusion of some additional sentences of the same reading topics. The recordings were done in small and quiet office rooms and a Zoom H2n portable recorder with five built-in microphones was used (Mid-Side recording mode). Most of the speakers of the Extended DigiSámi corpus did not consent to have their speech data publicly available (outside the University of Helsinki), and we respect their wish not to publish it.

Altogether, the Extended DigiSámi corpus contains material from 31 native North Sámi speakers and one fluent non-native. It is important to note that all North Sámi speakers are bilingual: the speakers from Guovdageaidnu and Kárášjohka were also native in Norwegian, while those from Ohcejohka, Avvil, and Anár were bilingual in Finnish and North Sámi.

In terms of the dialectal varieties, only the Guovdageaidnu areal variety belong to the Western dialect area according to the traditional analysis of North Sámi (see Sammallahti, 1998). The rest of the varieties belong traditionally to the Eastern dialect group of North Sámi (see Figure 1). In Guovdageaidnu and Kárášjohká, there are exceptionally large and active Sámi communities and North Sámi is spoken by the majority of the inhabitants.

The original Extended DigiSámi data were previously manually annotated and segmented on sentence level. For this paper, we further annotated the entire Extended DigiSámi speech data semi-automatically at the word and segment levels, using the WebMAUS Basic forced alignment tool (Kisler et al., 2017). Each word was also POS tagged. Additionally, the sentences were divided into phrases, following the natural syntactic structures of North Sámi, resulting in approximately three phrases per sentence. Note that this division was done on the textual material and only subsequently applied to the speech signal, therefore these phrases do not necessarily correspond to intonational phrases. Please refer to Appendix B for examples of the sentences and their division. All labels and segmentations were manually checked and corrected after any automatic procedures. At this point, however, we did not label for focus; there is unfortunately no research on North Sámi focus, and establishing a reliable framework for focus labeling would require its own, separate research contribution. The aim of our machine learning technique is to explore the data using as little manual annotation as possible.

Due to varying recording circumstances and certain voice quality characteristics of the speakers, we used a subset of the speech data described above. Unfortunately, one-third of the speakers (10 of 31) and a number of sentences (79) from the remaining speakers had to be excluded because of hoarse/creaking voice, hesitations, mistakes or severe background noise overlapping with normal speech frequencies. A sentence was discarded from the analyzed data if there were more than 10 errors in the F0 analysis (mostly octave jumps).

Two older speakers from Guovdageaidnu, excluded from the analysis, had a hoarse voice and many of hesitations, which made reliable pitch extraction impossible. The same applied for two cases (one middle-aged and one old) from Anár, Finland, where the voice of a speaker was so consistently hoarse or creaky that there were too many (more than 10) errors in the pitch extraction in nearly all sentences from that speaker. In these cases, the data would not have been reliable in terms of the prosodic analysis and a decision was made to exclude the speaker from the data. A number of speakers (six young speakers from Avvil, Finland) were excluded from the data because of a consistent buzzing noise in the critical speech frequencies which made it impossible to filter out the noise without compromising the actual speech signal.

The details (site, dialect, duration and age groups) of the remaining and analyzed data set are described in Table 1. Altogether, eight speakers from Finland and two speakers from Norway were excluded from the analysis. This resulted in an unbalanced amount of speakers from Finland (nine speakers) compared to Norway (12 speakers) but the difference between the amount of speech data in minutes (107 minutes vs. 114 minutes, respectively) was not considered to be remarkably different in a way that it would distort our analysis.

OPEN IN VIEWER

Table 1.

Description of the read speech data analyzed in this research. The thick line separates the Finnish and Norwegian sites. The total duration for the Finnish sites was 106:59 minutes and for the Norwegian sites 114:49 minutes. Age groups: Y = young (15–25); M = middle (25–50); and O = old (50–70).

Majority language	Site	Dialect	Speakers (Female)	Audio duration (minutes: seconds)	Age group distribution
Finnish	Anár	East	2 (2)	21:10	1Y 1O
Finnish	Avvil	East	1 (1)	11:32	1Y
Finnish	Ohcejohka	East	6 (1)	64:07	1Y 3M 2O
Norwegian	Guovdageaidnu	West	6 (4)	71:24	1Y 3M 2O
Norwegian	Kárášjohká	East	6 (4)	43:25	4Y 1M 2O
Total			21 (12)	221:48

5.2 Network architecture

The WaveNet (Oord et al., 2016) is a deep artificial neural network speech synthesis architecture that operates directly on the digital (pulse-code modulation) signal. In essence, the network is trained to output an individual signal sample using a preceding portion of the signal samples as input. As illustrated in Figure 3 (bottom right shaded areas) the input samples are processed by several stacked-up dilated convolutional layers. Every convolutional layer contains a number of convolutional filters whose output depends on values of several (two, in the case of this particular architecture) outputs from the previous layer. As shown in Figure 3, at each time-step, every output from the first hidden layer depends on two subsequent input samples. The following hidden layer, however, combines the values from the previous layers that are separated (dilated) by two time-steps, rather than subsequent values; each value thus depends on four input samples. The convolutional filters on the next layer are then dilated by four time-steps, and each output value is a function of eigth input samples.

OPEN IN VIEWER

Figure 3.

An architecture of the WaveNet synthesis system with the conditioning embeddings using prosodic signal.

This architecture facilitates an exponential increase of the length of receptive field (the previous portion of the input signal that determines the value of every given output sample) with every stacked convolutional layer, and provides a sort of parallel hierarchical analysis with more dilated layers capturing progressively longer-term dependencies in the signal.

In the present work we used a TensorFlow (Abadi et al., 2016) implementation of the WaveNet network architecture with two stacked-up sets each containing nine stacked layers with dilations incrementally doubled for each subsequent layer: 1, 2, 4, . . ., 256, 1, 2, 4, . . ., 256 (repeating several identical stacks of appropriately dilated convolutions does not increase the receptive field size, but is a common practice). This leads to the receptive field of length 1024 samples. See Appendix A for a technical description of the network used in this work.

The receptive field size determines the maximal temporal scope of dependency that can be learned by the network. As we are aiming to use our WaveNet implementation for a prosodic analysis, we want to be able to capture dependencies between events separated by at least several hundreds of milliseconds. While the receptive field size is a function of the number of hidden convolutional layers, its temporal scope obviously also depends on the sampling rate of the signal. Therefore, we trained the network on a low sample rate prosodic signal created from the high sample rate original speech signal.

5.3 Pre-processing and prosodic signal generation

As we aim at the analysis of the speech material in terms of prosody (rather than, e.g., segmental information), a new prosodic signal was generated for each sentence matching the original waveform in the F0 and energy envelope but containing no harmonics of the F0. This prosodic signal, designed to capture only the pitch and intensity of the original speech is sampled at a relatively low sampling rate of 800 Hz; this means that the 1024 samples of the receptive field covers 1.28 seconds of speech.

There are several reasons behind generating a completely new signal rather than using a downsampling filtering-based procedure. One was to avoid inclusion of speaker and site specific interference in the automatically analyzed material, such as background noise, room reverberation, etc. Another motivation is to recast the F0 contours of individual speakers to a common baseline (of 100 Hz median F0 for each speaker) while keeping the relative pitch movement intact (see below).

An alternative option satisfying both these criteria would be to train the network on paired low sampling rate F0 and intensity signals. While this approach is certainly viable and should be investigated in the future, in this work we opted for generating prosodic signals instead for two reasons: (a) the WaveNet network architecture has been primarily used and tested on “speech like” signals (intensity modulated carrier frequency signals); and (b) in our experience the architecture does not work well with two parallel signals.

The prosodic signal was created by the following procedure.

The F0 was extracted using a customized interactive Praat (Boersma & Weenink, 2016) script, allowing for manual checking and correcting octave jumps, and labeling creaky phonation, noise. and other artifacts. Interpolated contours were used for unvoiced intervals as well as for the intervals with identified creaky phonation. The resulting F0 contours were sampled at 800 Hz. Identically sampled (at 800 Hz) energy envelope e was also extracted from the signal by filtering a square of the original waveform signal filtered by a 25 milliseconds Hamming window.

In order to remove the pitch differences among speakers caused by natural physiological variability (leading to, e.g., differences between males and females) the pitch contours were modified so that the median F0 for each speaker equaled 100 Hz. The extracted F0 contours were first recast to semitone scale using the median F0 for each speaker (over the entire material) as a base frequency, and then new F0′ signals in Hz scale were recomputed using the base frequency of 100 Hz.

The envelope-modulated sinusoidal prosodic signal was then generated as

s n = e n sin ( 2 π s r ∑ i = 1 n F 0 ′ i ) ,

where e i and F 0 i denote the i th sample of the envelope and F0′ signals, respectively, and s r = 800 is the sampling rate ¹ .

As illustrated in Figure 4, without the pitch normalization described above, the signal s has the same F0 and energy envelope as the original, but contains no segmental or spectral information. Due to the pitch normalization, the resulting prosodic signals used here have F0 equal to F0′.

OPEN IN VIEWER

Figure 4.

Original audio waveform (in gray) superimposed with a sinusoidal prosodic signal (in black).

5.4 Embeddings

In addition to conditioning by the previous signal, the WaveNet architecture uses (local or global) conditioning to generate a signal with the required characteristics (Oord et al., 2016). The known characteristics of the signal are fed as an additional input to each dilated convolutional layer via embedding layers (see the network connections rendered in red in Figure 3), trained together with the other network components. The embedding layer learns to map a discrete set of control parameters (e.g., a set of individual words in the corpus) to real-number valued vectors that are directly used as an additional input to each convolutional layer (the matrix shown in the left portion of the shaded area in Figure 3). In the case of local conditioning used here, the appropriate embedding vector trained in a particular time is determined by a parallel signal containing the relevant control parameter value corresponding to the trained signal sample. In Figure 3, the different values of this parameter (corresponding to different words in the utterance) are schematically depicted by different coloring of waveform background and the corresponding coloring of the rows of the embedding matrix.

The main aim of the machine learning part of the presented method is to obtain vector representations of individual lexical items as they are produced by, on the one hand, the Sámi speakers from Finland and, on the other hand, by the speakers from Norway. Also, as frequent words might have been produced in multiple sentences and sentence positions, we want to at least to some extent account for possible differences that are attributable to sentence-level prosody.

With this aim, two local conditioning embeddings were implemented in the network architecture and trained in parallel: a word embedding; and a phrase embedding.

As stated in our methodological hypothesis (MH), the word embeddings are designed to capture the conditioning participating in the generation of the individual lexical items present in a corpus. Using the word-level annotation of the corpus, each sample of the prosodic signal was assigned a label corresponding to the word within which it occurred. For a given word in the corpus, one label was assigned to all renditions of the word by the speakers from Finland, and another label to the renditions by the speakers from Norway. The silent gaps between the words were assigned a separate “placeholder” label (illustrated in light-blue in Figure 3). Only the top 199 most frequent words in the corpus were assigned individual labels; the remaining words were grouped together under two common labels (one for the Sámi speakers from Finland, one for those from Norway). Altogether, 401 ( 1 + 2 × 199 + 2 ) conditioning vectors were trained.

The phrase embedding, referring to a part of an individual sentence in the text corpus, is designed to account for the prosodic variation associated with the position of the given portion of prosodic signal (corresponding to a word) within a particular phrase in a sentence. If the same lexical item occurred in multiple sentences (or parts thereof) its word embedding vector learned the representation of the common aspects of all these realizations, while the phrase embedding vector (different for different contexts) captures the differences associated with the given portion of the sentence.

The labels were assigned using the phrase annotations in the corpus; again separate labels assigned to “Finnish” and “Norwegian” Sámi renditions. The majority of the sentences (altogether 148 individual sentences) in the text materials were divided into approximately three shorter phrases. To follow the natural North Sámi syntactic patterns and to keep the phrase length as comparable as possible, shorter sentences were not divided into phrases at all. Altogether, 284 different phrases were identified in the corpus, resulting in 569 phrase conditioning vectors ( 2 × 284 + 1 for the silent gaps between phrases).

Please note that we primarily focus on the word embedding and use the phrase embeddings to automatically account for the occurrences of the given word in different sentences and sentence positions. This decision does not mean that we are not aware of the possibility of prosodic phenomena at phrase-level or beyond. The scope of phenomena that can be investigated using our chosen approach is, however, limited by the length of the receptive field (a bit over one second in the present implementation).

We provide the prosodic signals used in this work, the corresponding word and phrase labeling files as well as the trained embedding as Online Supplementary Material for this article.

To summarize, the task of the network is to learn to predict the next sample of the prosodic signal, given its phrase and word identity, as well as the speaker’s country of origin. In order to generate an appropriate prosodic signal, the fully trained network would receive two parallel sequences of (appropriately aligned) word and phrase labels; as there is no temporal modeling attempted in the present implementation, the word/phrase sequences would need to be of appropriate length. The network would generate a “Norwegian” or “Finnish” signal, depending on the labels (mixing the languages would also be possible by feeding a “Finnish” word label alongside a “Norwegian” phrase label, or vice versa).

Given these input labels (and a default “starting” signal sequence of the receptive field length), the network would generate the signal in an auto-regressive manner, using the previously generated samples as an input.

The WaveNet implementation used in this paper was, however, not designed and tuned to yield a high-quality synthesized prosodic signal. Instead, we focus on the learned word representations derivable from the trained network weights. In particular, we analyze the differences between embeddings for individual words as uttered by the “Finnish” and “Norwegian” Sámi speakers.

5.5 Training procedure

All data were used to train the network, as a training set. Obviously, this decision means that the network may overfit the data (fail to generalize to other similar tasks), and that its performance cannot be subsequently tested on a previously unseen test set. This approach would be clearly detrimental if the task was to train a speech synthesizer that can generate signals for previously unseen sequences of words. The aim of this work is, however, to capture the statistical properties of the entire corpus material. Moreover, what is subsequently analyzed is not a generated signal but part of the trained network itself, namely the word embedding layer, that does not change with the input to the network after the training.

A validation set comprised a subset of 20% of the data and was used for stopping the training process: training was stopped after 300 epochs, or earlier if the loss calculated on the validation sub-set did not improve for 25 subsequent epochs.

The training was repeated 10 times, and the word embeddings were extracted after each training run.

6.1 Word embedding distance and comparison of the models

The analysis was limited to the words with at least 17 separate occurrences for each language group in the corpus; 70 words fulfilled this requirement. The number 17 was chosen rather arbitrarily, but it yields a sufficient number of tokens for the analysis and guarantees enough repetitions for each token. Using other cut-off numbers in the vicinity of 17 does not qualitatively influence the subsequent results.

To address our typological hypothesis (TH) and in order to obtain a measure of differences between the prosodic realizations of these words as captured by the trained models, for each word we calculated the distance between the embedding vector for the word for the “Finnish” Sámi speakers from the embedding for the “Norwegian” Sámi speakers. First, each trained embedding matrix was “spherized” by first subtracting the mean from the every embedding vector and then by dividing every column in the resulting matrix by its standard deviation. Second, as it is likely that some of the embedding vector dimensions did not capture any relevant information, we transformed the spherized vectors using a principal component analysis and used only the first 32 dimensions (explaining about 75% of variance) in subsequent analysis. Spherization and principal component analysis were performed using MATLAB.

Figure 5 shows the resulting 32-dimensional vectors of one model, recast to a two-dimensional space using the Uniform Manifold Approximation and Projection (UMAP) (McInnes et al., 2020) method. We observe that despite the lack of spectral information in the signal, the phonotactically similar words are generally clustered together in the embedding space. In addition, the Norwegian and Finnish variants of the words are close to each other, but with varying distance. Direction between variants does show some pattern, but there is no general trend between language variants, suggesting that our signal processing procedure successfully mitigated the possible group differences between “Norwegian” and “Finnish” speakers, unrelated to our research question, such as, for example, difference gender/age distributions, potential differences in recording conditions, etc. Note that the distances in this two-dimensional projection do not directly correspond to the actual embedding distances discussed below.

OPEN IN VIEWER

Figure 5.

Uniform Manifold Approximation and Projection of embeddings of 30 closest word pairs: red, Norwegian Sámi; and blue, Finnish Sámi.See the online article for the color version of this figure. See the online article for the color version of this figure.

For each word, a Euclidean distance between the pair of 32-dimensional embedding vectors—one for each majority language group—was calculated. In order to verify that the trained models yielded reasonably similar word distance measures, we calculated pairwise Pearson correlation coefficients comparing the word distance vectors for different models. The correlations ranged between 0.80 and 0.90, with a mean of 0.85. Consequently, we averaged the word distances for each word across all 10 trained models; these mean distances will be used in the analysis below.

The barplot in Figure 6 shows the embedding distances between the Finnish and Norwegian realizations for the 70 most frequent words in the corpus. Note that for the top 10 words (the portion with the gray background from “olu” to “Ruota”), the distances are somewhat greater than for the remaining words (see the “offset” between “Ruota” and “ledje”). An inspection of the speech material revealed that this is probably caused by frequent hesitations and corrections within these words in the corpus; therefore, these words were excluded from the subsequent analysis.

OPEN IN VIEWER

Figure 6.

Average embedding distances for the most frequent words in the corpus. Coloring reflects the part-of-speech category of each word. Note that the inflections of the copula (lea, leat, ledje) are altogether more frequent in the text than the other verbs which might also reflect in the prosodic renditions of these words.

The bar colors in Figure 6 indicate the POS categories of individual words. Note that the “green” colored words (nouns, adjectives, and numerals) primarily occupy the lower part of Figure 6, while verbs, adverbs, adpositions, and pronouns tend to occur in the upper portion. This observation is supported by the mean embedding distances for different POS categories listed as part of the legend in Figure 6.

6.2 Comparing word distance with F0 characteristics

Can the differences between the Finnish and Norwegian Sámi renditions captured by the embedding distances be associated with some particular characteristics of the prosodic signals (as also formulated in our methodological hypothesis (MH))? In order to address this question and the MH we extracted two F0-related measures from the original corpus, namely the mean F0 range and mean F0 standard deviation (SD) over the Finnish and Norwegian renditions of each analyzed word.

More precisely, given a word, corresponding F0′ (the median-normalized pitch) values were extracted for every occurrence. Then, standard deviation of these values was calculated, and these standard deviations were averaged across all occurrences for Finnish speakers and all occurrences for Norwegians, yielding the group specific mean F0 SD for the word. Similarly, the differences between 90th and 10th percentiles of the F0′ values were used to calculate the group-specific mean F0 ranges for each word.

Finally, the difference of these two measures between Finnish and Norwegian speakers were calculated. F0 SD distance for a word is the Finnish mean F0 SD minus Norwegian mean F0 SD for the word. Similarly, F0 range distance is the Finnish mean F0 range minus Norwegian mean F0 range for the same word.

When all 60 words are taken together, neither of these distance measures significantly correlates with the embedding distance: for the SD distance, the Pearson correlation is 0.16 ( p = 0 . 21 ), and for the range distance the correlation is 0.05 ( p = 0 . 72 ).

The situation is somewhat different when the words are grouped by their POS categories. Table 2 shows correlations between the SD distance and range distance, respectively, and the embedding distance for different types of words. While in most cases the correlation coefficients failed to reach significance (presumably due to relatively low counts; column n in Table 2), the correlations reveal an interesting pattern.

OPEN IN VIEWER

Table 2.

Correlations between the embedding distances and the acoustic distance measures, for different part-of-speech (POS) categories.

POS	n	Mean fundamental frequency (F0) standard deviation		Mean F0 range
		Coefficient	p value	Coefficient	p value
Nouns	28	0.38	0.045	0.25	0.209
Adjectives	4	0.28	0.727	−0.07	0.932
Conjunctions	7	0.46	0.299	0.48	0.275
Verbs	7	−0.53	0.220	−0.57	0.186
Adverbs	7	−0.68	0.095	−0.64	0.119
Pronouns	3	−0.63	0.564	−0.70	0.505
“Noun” group	34	0.43	0.012	0.23	0.196
“Verb” group	19	−0.57	0.012	−0.57	0.010

The coefficients and the respective p values marked in bold are considered as statistically significant (p < 0.05).

The correlations are positive for nouns, adjectives, and conjunctions, indicating that the greater embedding distance is associated with greater F0 movement (in terms of SD and range) for Finnish renditions than for Norwegian ones for these categories. For the remaining POS types, the correlations are negative, meaning that for the most distant words (in terms of embeddings) the pitch excursions are larger in the words uttered by the Sámi speakers from Norway compared to the ones from Finland.

The bottom two rows in Table 2 list the correlations for two groups of words: the first, “noun” group containing nouns, adjectives, and numerals; and the second, “verb” group containing verbs, adverbs, adpositions, and pronouns. With the exception of the correlation between range and embedding distances for the “noun” group, the correlation coefficients are significant for the grouped word categories. Note that we use the group names “noun group” and “verb group” as shorthand names for the groups since they mostly contain nouns or verbs and words that have similar syntactical roles in North Sámi, respectively.

This observation is illustrated in Figures 7 and 8 showing the relationships between the SD distance and range distance, respectively, and the embedding distance. The coloring of the words indicates their belonging to the two POS groups and the linear fits show the nature of the relationships between the distance measures obtained using the WaveNet embeddings and more traditional acoustic measures.

OPEN IN VIEWER

Figure 7.

Embedding distance of words as a function of the mean fundamental frequency standard deviation distance. The “noun” group in green, the “verb” group in orange, the conjunctions in gray.

OPEN IN VIEWER

Figure 8.

Embedding distance of words as a function of the mean F0 range distance. The “noun” group in green, the “verb” group in orange, and the conjunctions in gray.

6.3 Comparing F0 contours for individual words

Figures 5–8 can be used to inform us about which individual words might show the clearest differences between the two majority language groups. Of course, the greatest embedding distance does not automatically mean that the given word provides the best example of prosodic phenomena of interest. The link between the quantified intonation phenomena and embeddings is indirect: the differences between the “Finnish” and “Norwegian” renderings are, presumably, signaled by relatively great distances between the respective embeddings, and these, in turn, correlate with SD distance and range distance. The fact that these measures correlate does not necessarily mean that every word yielding a great embedding distance will also manifest the relevant phenomena in an interpretable way.

In what follows, we present a more detailed comparative analysis of several words (from the two POS category groups) with a considerable embedding distance (circled in Figures 7 and 8) and compare them with words that yielded the smaller embedding distances (rectangles in Figures 7 and 8). The words that were selected illustrate the link between greater F0 range/SD distances and the embedding distance in the most straightforwardly interpretable way.

For selected words we calculated the mean F0 contours and standard deviation bands, separately for the ”Norwegian” and ”Finnish” North Sámi speakers. Using the segment-level annotations, mean segment durations were calculated for a given word’s renditions by Sámi speakers from Norway and those from Finland. These average segment durations were used for time-normalization, with every segment divided into seven equally spaced intervals. Mean F0 values and standard deviations at the corresponding interval boundaries were calculated using the speaker-normalized F0′ contours, and time-warped to the corresponding points in the average segment intervals. The resulting time-normalized and speaker-normalized pitch contours and standard deviation bands are plotted in Figures 9–11; the blue and red corresponding to ”Norwegian” and ”Finnish” Sámi participants, respectively; the vertical lines mark (average) segment boundaries; and the white circles mark the time-normalization points. For all of the selected example words in Figures 9–11, the y-axis shows the normalized mean F0 in Hz, and the x-axis indicates the normalized time in seconds. The sentence contexts of the example words are shown in Appendix B. Senrence contexts are also shown in Appendix B.

OPEN IN VIEWER

Figure 9.

Comparison of mean fundamental frequency (F0) for words from the “verb” group (verbs, adverbs and adpositions): red, Norwegian Sámi; and blue, Finnish Sámi. Shaded region indicates standard deviation. The green region emphasizes the differences in the mean F0 between the two averaged renditions of the words. The word meanings, part-of-speech labels and expected placement of lexical stress (in bold): d á lá ái gái—“until nowadays” (adverb); j o htá—“to pass/carry” (verb); and ea rá—“other” (adverb). The F0 values are plotted in Hz scale and created using the range and time normalized F0 contours.See the online article for the color version of this figure. See the online article for the color version of this figure.

OPEN IN VIEWER

Figure 10.

Comparison of mean fundamental frequency (F0) for words from the “noun” group (nouns and adjectives): red, Norwegian Sámi; and blue, Finnish Sámi. Shaded region indicates standard deviation. The green region emphasizes the differences in the mean F0 between the two averaged renditions of the words. The word meanings, part-of-speech labels and expected placement of lexical stress (in bold): b ea ltta—“a belt” (noun); g á vttit–“the Sámi costumes” (plural, noun); v e jolaš—“possible” (adjective); and d e hálaš—“important” (adjective). The F0 values are plotted in Hz scale and created using the range and time normalized F0 contours.See the online article for the color version of this figure. See the online article for the color version of this figure.

OPEN IN VIEWER

Figure 11.

Comparison of mean fundamental frequency (F0) for words with similar prosodic realizations: red, Norwegian Sámi; and blue, Finnish Sámi. Shaded region indicates standard deviation. The green region emphasizes the differences in the mean F0 between the two averaged renditions of the words. The word meanings, part-of-speech labels and expected placement of lexical stress (in bold): l ea t—“to be” (copula, verb); g á kti—“a Sámi costume” (noun); s á miid—“the Sámi people” (plural genitive, noun); and s á megiela —“the Sámi language” (singular genitive, noun). The F0 values are plotted in Hz scale and created using the range and time normalized F0 contours.See the online article for the color version of this figure. See the online article for the color version of this figure.

Figure 9 shows examples of mean F0 contours (time aligned and speaker-normalized) from the “verb” group including verbs, adverbs, and adpositions. In this category, there were certain different patterns in the Norwegian and Finnish Sámi areal varieties. All of the four example words are disyllabic, and in the Norwegian variety, the highest F0 peaks of the word tend to occur within the first syllable. Contrastively, in the Finnish variety, there is a valley in the first syllable, and the F0 rises towards the second syllable and towards the beginning of the next word, likely a noun. Generally, it can be noted that for the Norwegian variety, there is more movement in F0 compared to the Finnish one. The two adverbs plotted in the top panels of Figure 9, dálá and áigái, forming a phrase dálá áigái (meaning “until nowadays”); note the U-shaped overall F0 contour for the Finnish rendition and the more complex Norwegian pattern with a potential peak in the first syllable of the second word.

In the contours in Figure 10, including nouns and adjectives, the F0 of the example words have more movement (or greater range) for the Finnish variety than the Norwegian variety, particularly in the latter part of the words. In Finnish, the first syllable has always the primary stress while the second or third syllable bears a secondary stress (Suomi & Ylitalo, 2004; Suomi et al., 2003, 2008, pp. 75–78). If nouns are more prominent in terms of F0 range or movement (two peaks), the latter F0 peak could potentially correspond to the secondary stress in Finnish North Sámi renditions. In these examples, the timing of the peaks are different in the varieties, especially in the disyllabic words (bealtta and gávttit). Interestingly, both varieties have the highest peak in the beginning of almost all of these examples, differing from the examples in Figure 9. As a word of caution, it has to be noted that two F0 peaks in the average contours do not necessarily indicate two peaks in individual renditions; they may also result from averaging over two distinct patterns (one with an early and one with a late peak) for the particular example.

The differences between the example words in Figures 9 and 10 might look relatively small, but when comparing them with the similar ones in Figure 11, certain patterns can be observed. As the language is the same, with speakers from different areas, radical differences are not expected. What we are looking for are small prosodic differences that could be traced back to the prosodic characteristics of the majority languages. The translations of the sample words in Figures 9-11 as well as their sentence contexts are also shown in Appendix B. In Figure 11, there are examples of some of the most similar words between the two varieties. The F0 contours follow mostly the same directions in not only the disyllabic but also in the 4-syllabic example word. The unifying factor for three of these words is that they are connected to the Sámi culture and are presumably used frequently by the speakers, regardless of the areal variety. One of the examples, leat, is a copula, which means it is a very frequent word and it often gets considerably segmentally reduced (e.g., from /leaht/ to /la/) and thus its prosodic rendition might assimilate more to the prosodic patterns of the surrounding words. When considering the duration of all example words (the x-axis), it seems that the Norwegian variety is generally slightly longer in all of the example words.