ИСТИНА

In our project some perception experiments were performed to establish the number of prosodic boundary strength (PBS) levels that unprofessional auditors can reliably detect in a spoken Russian text. This paper deals with the statistical analysis of these experiments’ data. Nonparametric and nonnumeric statistical methods were applied. All computations were realized by means of statistical program package STATISTICA. The introduction provides a brief description of the background and discusses concept of PBS level number detection using the consistency of the auditors group as concordance (agreement) of their break indices (BI) labeling of the same text. In the next part of the paper different measures of consistency and algorithms of coordination procedures for auditors’ group are discussed. An array of twenty break indices labeling of the 470-words’-spaces-long prosaic Russian text performed by 19 unprofessional auditors and one highly qualified phonetician − participants in the experiment is considered as an example. By means of iterative coordination procedure of step by step removal of the “worst” auditors, i.e. those ones, whose BI labeling have the maximum deviation from current coordinate labeling, the resulting BI labeling is obtained as termwise median mean for labeled texts of most consistent resulting auditors’ group. Similar break indices labels could be obtained as a centroid of the most dense claster whereas the formal clastering could be received by modified K-means method with cross-validation (STATISTICA, Claster Analysis and DATA MINING sections). In the last part of the paper the dependence between the PBS levels in resulting break Indices labeling and pause localization and their timing in the same spoken text is investigated. The most probable values and confidence intervals for pause duration corresponding to various PBS levels are estimated.