The course consists of lectures, three practical laboratory exercises with assignments and a written essay on a chosen title. The essay will be presented orally during a closing seminar.
Included topics:
- algorithms for training, recognition and adaptation to speaker and transmission channel, mainly based on Hidden Markov Models (HMMs) and Deep Neural Networks (DNNs)
- methods for reducing the sensitivity to external noise and distortion
- probability theory
- signal processing and feature extraction
- acoustic modelling of static and time-varying spectral properties of speech
- statistic modelling of language in spontaneous speech and written text
- search strategies – basic methods and algorithms for large vocabularies
- specific analysis and decision techniques for speaker recognition.
The laboratory exercisea are intended to give practical experience of designing different aspects of a speech recognition application. It consists of the implementation of functions given prototypes and on testing those functions with real speech data.