Sixth European Conference on Speech Communication and Technology
This paper presents an experimental speech recognition system used to transcribe a weather forecast broadcasted by the Czech radio. The system is based on the HMM with mixture Gaussian continuous densities and is designed as a speaker independent. To overcome very sparse training data various language models supported by smoothing of model parameters based on the leaving-one-out technique, discounting and backing-off approach were tested. The results of recognition experiments are discussed in the paper.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Müller, Ludek / Psutka, Josef (1999): "Using various language model smoothing techniques for the transcription of a weather forecast broadcasted by the czech radio", In EUROSPEECH'99, 1783-1786.