Vowel Processing in Cluttered Auditory Scenes

Beatrijs van Meerveld

Research output: ThesisThesis fully internal (DIV)

466 Downloads (Pure)

Abstract

Problems with speech processing are associated with the poor representation of speech and a late use of knowledge to disambiguate the input sounds. The first goal of the current work is to improve acoustical features. We selected the local energetic components from a harmonic complex that correlate to speech articulations. We show that these extractions facilitate noise-robustness. The usefulness for speech recognition could not be explicitly tested with modern ASR techniques because the number of extractions varies which does not fit current systems for ASR.

We assume that humans, in contrast to ASR techniques, are able to process a variable number of input features. Therefore, we investigated, as a second goal, the effect of knowledge and expectation on human perception of vowels. The research paradigm where speech in noise can be seen as speech with additional, irrelevant features provides new ways to understanding and modelling speech processing. We performed two perceptual experiments. In two experiments we presented manipulated Dutch spoken vowels. We conclude that the human perceptual system can flexibly handle the varying number of extractions. Knowledge of sounds in a language leads to integration of part of the features that together form a coherent percept.
Original languageEnglish
QualificationDoctor of Philosophy
Awarding Institution
  • University of Groningen
Supervisors/Advisors
  • Schomaker, Lambert, Supervisor
  • Baskent, Deniz, Supervisor
  • Andringa, Tjeerd, Co-supervisor
Award date25-Nov-2016
Place of Publication[Groningen]
Publisher
Print ISBNs978-90-367-8755-0
Electronic ISBNs978-90-367-8754-3
Publication statusPublished - 2016

Cite this