Interspeech'2005 - Eurospeech
In the paper a new automatic intonation recognizer for the Polish language is presented. The recognizer design combines Machine Learning and expert knowledge techniques. Machine Learning is used in pitch stylization (Artificial Neural Network), speech alignment (external design based on Hidden Markov Model) and intonation decoding (Hidden Markov Model). Expert knowledge drives phonemization, syllabification, lexical accentuation and lemmatization. In the recognizer, a recently available intonation annotation system for Polish is used. The intonation annotation system and the related expert knowledge allowed for substantial reduction in the number of designed HMM decoder parameters. Software integration problems emerging from the number of modules comprising the recognizer are approached using an original software environment for speech processing. The environment has data-centric message-driven Blackboard-like architecture with an annotation graph as a shared memory.
Bibliographic reference. Wypych, Mikolaj (2005): "An automatic intonation recognizer for the Polish language based on machine learning and expert knowledge", In INTERSPEECH-2005, 3305-3308.