Linguistic information is hard-coded in speech signal. By analysing specific acoustic properties, such as vowel formants and fundamental frequency, acoustic models of speech production aim to elicit this information. I summarise evidence from my research on speech processing and argue that these acoustic models provide only incomplete spectral description and under-represent interactions between acoustic properties. Consequently, they do not do justice to the complex linguistic information encoded in speech.
I then propose a model for speech processing based on parameterised resonant signal elements and an algorithm that analyses vowel samples based on the proposed model. The algorithm provides a rich description of any given segmented vowel sample by using a large number of resonant elements with parameters that are chosen to accurately capture the time-frequency structure of the vowel. The parameters are then used to calculate probabilities. An application of the model successfully classifies vowels, stress, and speech variety. This model is an improvement over methods that only use a small number of formants to describe vowels, has the potential to be used in automatic speech recognition, and is promising for use in applications of forensic linguistics, and speech pathology. Finally, I discuss an ongoing work that aims to extend the model for the analysis of prosody.
@Language and Probability: The CLASP Inauguration Workshop
The Centre for Linguistic Theory and Studies in Probability (CLASP) at the Department of Philosophy, Linguistics and Theory of Science, University of Gothenburg hosted a one-day workshop to inaugurate the Centre on Thursday 27th August 2015 at Gothia Towers.
The workshop focused on core areas of CLASP’s research mission. It featured short talks by representatives from the university, the Swedish Research Council, CLASP researchers, and members of CLASP’s international Scientific Advisory Committee.
The workshop was collocated with Semdial 2015 – goDIAL.