Year of Graduation
Speech Rate Study Based on Data from Spoken Corpora of Languages of Russia
Fundamental and Computational Linguistics
After the studies on individual differences in speakers’ speech rates, more general studies on the speech tempo occurred. They were focused on the investigation of speech rate of a language as a whole and typological comparison of languages by this parameter. All these works have a non-formulated assumption that the speech rate is a non-observable language-specific feature that varies from language to language. From this perspective, individual speech rates are somehow derived from this language specific feature. The opposite point of view is that the observed variability in individual speech rate is only partially affected (or even not affected at all) by language itself, and the main factors that influence variability are speaker-specific (age, sex, etc) and discourse-specific (length of the phrase in syllables, emotionality of speech etc). This work explores the relevance of the speech rate as a typological language-specific parameter. The hypothesis is that a language as a parameter does not affect a speakers’ individual speech rate. To testify the hypothesis, I study individual speech rates of speakers of Russian, Azeri, Bashkir, Beserman dialect of Udmurt, and Chukchi languages. The analysis is based on the data from spoken corpora of the languages. With this data, I have designed several statistical Multilevel Mixed-Effects Models, representing different assumptions regarding the structure of the data and the importance of factors. The Multilevel Mixed-Effects Models were used because of the lack of independence between observations, as one observation in the sample equals to one utterance and there are multiple observations from one speaker. The independence of observations in one of the fundamental assumptions of the regression analysis. The Multilevel Mixed-Effects Models allow using regression analysis in a situation of independence violation. There were six models, grouped in pairs. Each pair had a different combination of population-level and group-level factors. Within a pair, models differed only by the parameter of presence/absence of language group-level factor. In accordance with the hypothesis, it was expected, that models without language group-level factor will describe the data at least not worse than models with this factor. The results of the statistical comparison of the models do not support the advanced hypothesis: the model, containing language factor described the data significantly better, than the models not containing it. On the other hand, the value of the main effect (the length of an utterance in syllables) estimate, provided for this model, is below the linguistical significance for the most part of the data. Therefore, these models have to be improved by an additional amount of data, that may strengthen the already discovered dependency or reveal new ones. Until then, I do not consider it possible to definitively confirm or disprove the hypothesis of the irrelevance of a language speech rate as a cross-linguistic parameter. Nevertheless, the present work provides an algorithm to study various linguistic parameters in terms of their influence on individual speaker’s characteristics. Using it, it is possible to provide strong evidence for or against the linguistic generalizations of different levels.