Lourdes Aguilar

Hiatus and diphthong: Acoustic cues and speech situation differences

Abstract The aim of this study is to determine the acoustic properties of hiatuses (vowel-vowel sequences) and diphthongs (glide-vowel sequences) in Spanish and to observe how these properties are modified depending on communicative factors. To do this, two groups of data were used: speech samples gathered from conversations between two speakers participating in the execution of a map task, in which the corpus items corresponded to toponyms, and the reading of the same sequences at a normal speaking rate. The comparison was done phonetically and phonologically: first, diphthongs and hiatuses were analyzed acoustically, studying their duration and spectral dynamics, and later, an inventory of diphthongizations and monophthongizations was made. Results show that hiatuses and diphthongs differ in the temporal and frequential domain: hiatuses have a longer duration and a greater degree of curvature in the F2 trajectory than diphthongs. Differences between the two categories (hiatus and diphthong) exist in both communicative situations, although changes within each category due to the speech situation were also observed: sequences from the map task are shorter and the degree of curvature of their formant trajectories is lower than for the reading task. We have also found that vowel-vowel and glide-vowel sequences behave differently in the way they are phonetically reduced: a reduction axis can be drawn in which hiatuses become diphthongs, and diphthongs vowels. It is concluded that hiatus and diphthong are two phonetic categories which can be described on the basis of their acoustic characteristics and are subject, like any other phonetic category, to modifications due to a change in the communicative situation.

Analysis of inter-transcriber consistency in the Cat_ToBI prosodic labeling system

A set of tools to analyze inconsistencies observed in a Cat_ToBI labeling experiment are presented. We formalize and use the metrics that are commonly used in inconsistency tests. The metrics are systematically applied to analyze the robustness of every symbol and every pair of transcribers. The results reveal agreement rates for this study that are comparable to previous ToBI inter-reliability tests. The inter-transcriber confusion rates are transformed into distance matrices to use multidimensional scaling for visualizing the confusion between the different ToBI symbols and the disagreement between the raters. Potential different labeling criteria are identified and subsets of symbols that are candidates to be fused are proposed.

Analysis of inconsistencies in cross-lingual automatic ToBI tonal accent labeling

This paper presents an experimental study on how corpus-based automatic prosodic information labeling can be transferred from a source language to a different target language. Tone accent identification models trained for Spanish, using the ESMA corpus, are used to automatically assign tonal accent ToBI labels on the (English) Boston Radio news corpus, and vice versa. Using just local raw prosodic acoustic features, we got about 75% correct annotation rates, which provides a good starting point to speed up automatic prosodic labeling of new unlabeled corpora. Despite the different ranges and relevance of inter corpora acoustic input features, the contrasting of the results with respect to manual labeling profiles indicate the potential capabilities of the procedure.

Cross-lingual English Spanish tonal accent labeling using decision trees and neural networks

In this paper we present an experimental study on how corpus-based automatic prosodic information labeling can be transferred from a source language to a different target language. The Spanish ESMA corpus is used to train models for the identification of the prominent words. Then, the models are used to identify the accented words of the English Boston University Radio News Corpus (BURNC). The inverse process (training the models with English data and testing with the Spanish corpus) is also contrasted with the results obtained in the conventional scenario: training and testing using the same corpus. We got up to 82.7% correct annotation rates in cross-lingual experiments, which contrast slightly with the accuracy obtained in a mono-lingual single speaker scenarios (86.6% for Spanish and 80.5% for English). Speaker independent monolingual recognition experiments have been also performed with the BURNC corpus, leading to cross-speakers results that go from 69.3% to 84.2% recognition rates. As these results are comparable to the ones obtained in the cross-lingual scenario we conclude that the new approach we defend has to face up with similar challenges as the ones presented in speaker independent scenarios.