Vowel and Consonant Quantity in Southern German varieties

The highly successful collaboration between the IPS (Munich), the ARI (Vienna), and the Phonetics Laboratory (Zurich) will be continued in this renewal proposal which builds upon the large database and the results obtained during the initial project. We now explore issues related to the trajectory of a sound change from its origin to its spread, with foci on (1) the voicing contrast, (2) the experimental testing of specific hypotheses formulated in sound change theories, and (3) the exploration of potential spin-offs from basic research to speech technology. The present proposal embraces the opportunity to examine for the first-time gradual apparent- time changes in the hierarchy of acoustic cues to the phonemic voicing contrast during the process of phonological change while considering potential effects of regional intonation differences. The main aim is to contribute to a better understanding of the time course of phonological change. The proposal subsumes four specific aims: (1) to examine the development of trading relations between acoustic cues in diachronic change; (2) to further investigate how contact and interaction trigger sound change by means of agent-based modelling; (3) to provide experimental evidence for the sound change in progress to be one of lexical diffusion; and (4) to expand the analyses to dialect synthesis. The innovation of the renewal proposal is in the combination of phonetic analyses of speech production and perception data in the testing of linguistic theories, the application of computational methods to large-scale, cross-linguistic, apparent-time data (including data from older speakers and children), and the consideration of potential applications for speech technology. The proposed project links acoustic, perceptual, and computational methods for studying the production- perception-relation in speakers of different ages developed in Munich with methods from speech technology established at the ARI for analyses of Austrian regional varieties as well as with typology-based methods for quantifying Swiss varieties established in Zurich. In the long-term the collaboration aims at the development of a model of linguistic (in)stability that integrates sociolinguistic and experimental phonetic data.

In the second part of the project, we investigated at the use of the project data in speech synthesis and speech technology, as well as the phonetic analysis of the data. In an acoustic model for speech synthesis, which describes the relationship between the phonetic symbols and the spectral features, a technique of adaptation to speakers was applied to improve the synthesis of dialect speakers. An average voice was created from the data of Viennese standard and Viennese dialect speakers, and then adapted to the desired speakers by training an artificial neural network. In a cooperation with the University of Bilbao, a speech synthesis system for Spanish was also expanded to include a neural network that describes the relationship between the spectral features and the acoustic waveform. We were also able to show how speaker and dialect embeddings can be used to generate regional standards for dialect speakers or dialect speech with the voice of a standard speaker using a state-of-the-art neural network model. We also investigated the use of standard grapheme-to-phoneme conversion for dialect speech synthesis. In the area of speech recognition, i.e. the conversion of spoken language into text, we were able to show how automatic speech recognition for Austrian German can be improved by speaker augmentation using neural network-based speech synthesis. New speakers are generated using speech synthesis, which are then used to generate more training data for the speech recognition model. In the area of speech technology used in phonetics, we have worked on denoising historical data (data from Anton Pfalz). This technology should make it possible to subject historical recordings to phonetic analysis, for which it is necessary to have fine spectral details in the data. We have also investigated the influence of denoising methods on the automatic segmentation of acoustic speech data using the WebMAUS tool, which is used in the field of acoustic phonetics for automatic segmentation. In the area of phonetic analysis, we were able to show in a simulation how communication between speakers of German from Austria and Germany can change the sound model of the speakers, depending on the number of speakers in the respective group. For the standard varieties in Germany and Austria, a gradual transition was found, but this is perceived with a distortion, depending on the variety of the listeners. For the Viennese sociolects, we were able to show that they are influenced in a complex way by gender, age, and education. For Austrian dialects from Innervillgraten, Bad Goisern, and Vienna, we investigated the connection between continuous transitions in these varieties and the 2-competence model of linguistics.