Polynomial Modeling of Child and Adult Intonation in German Spontaneous Speech

Abstract

In a data set of 291 spontaneous utterances from German 5-year-olds, 7-year-olds and adults, nuclear pitch contours were labeled manually using the GToBI annotation system. Ten different contour types were identified. The fundamental frequency (F0) of these contours was modeled using third-order orthogonal polynomials, following an approach similar to the one Grabe, Kochanski, and Coleman (2007) used for English. Statistical analyses showed that all but one contour pair differed significantly from each other in at least one of the four coefficients. This demonstrates that polynomial modeling can provide quantitative empirical support for phonological labels in unscripted speech, and for languages other than English. Furthermore, polynomial expressions can be used to derive the alignment of tonal targets relative to the syllable structure, making polynomial modeling more accessible to the phonological research community. Finally, within-contour comparisons of the three age groups showed that for children, the magnitude of the higher coefficients is lower, suggesting that they are not yet able to modulate their pitch as fast as adults.

The Future of Prosody: It's about Time

“…Matching 2 nd degree polynomial contours are superimposed in two time domains in order to mark the contour trends (for polynomial fitting of intonation curves cf. [65], [66] and contributions to [67]. Two interpausal Units (IPUs) in the question show an overall falling-rising contour, the answer shows an overall falling contour, and the global contour over both question and answer follows a holistic rise-fall pattern.…”

“…also Figure4) shows a question-answer sequence from a broadcast interview (AixMARSEC J0104G). Matching 2 nd degree polynomial contours are superimposed in two time domains in order to mark the contour trends (for polynomial fitting of intonation curves cf [65],[66]. and contributions to[67].…”

Description of F0 contours with Legendre polynomials

“…This article presents a method for the description of melodic patterns with Legendre polynomials, which provide a quantification that is more linguistically meaningful than conventional polynomial approximations. This approach was already used for the analysis of British nuclear tones (Grabe et al, 2007) and it was applied also to German (de Ruiter, 2011) and Czech (Volín et al, 2017). Furthermore, the research on Czech has exploited Legendre coefficients in the field of automatic speech processing, where they were shown to effectively parameterize the nuclear patterns and improve the prosody of the TTS synthesis (Matura & Jůzová, 2018).…”

Low Resource Automatic Intonation Classification Using Gated Recurrent Unit (GRU) Networks Pre-Trained with Synthesized Pitch Patterns

“…For example, the nuclear tone in Glide-up and Take-off are the same, but these classes are discriminated based on the entire tone pattern. Most of the existing works have studied the variations of intonation among different nativities [13][14][15][16][17][18] and variations of BE intonation across the na-…”