Pattern matching based speech recognizers allow the user to define or expand the recognition vocabulary by simply speaking the words a few times. From which language these words are is completely irrelevant. However, such recognizers are much less robust against background noise than modern statistical recognizers. In these recognizers the phonemes of a language are represented by hidden Markov models (HMM). Consequently, these recognizers are not language-independent.
The aim of this project is a new type of speech recognizer that combines the simple and language-independent definition of the recognition vocabulary with the high robustness of HMM-based recognizers. The fundamentals of this project come from the recently completed thesis [Ger11].
Supported by: This project is mainly supported by KTI.