00000ctm a22000003a 4500 UP-99796217608810423 Buklod 20090505110425.0 m |o d | ta 090505s xx d r |||| u| (iLib)UPD-00084797013 DENGII eng LG 993.5 2009 E64 A46 Almonte, Reginald M. Incorporating natural language processing techniques in Filipino speech recognition Reginald M. Almonte, Michael Gringo Angelo R. Bayona, Patrick Simon T. Corrales. 2009 119 leaves ill. 1 computer laser optical disc (4 3/4 in.) Thesis (B.S., EEE)--University of the Philippines, Diliman. Previous projects in the UP Digital Signal Processing Laboratory have dealt with developing a system to recognize Filipino speech and convert it to text. Researchers, however, have observed that most of the word sequences generated the system do not resemble the natural form of the language. In this project, natural language processing (NLP) techniques were incorporated in the speech recognition system for continuous speech and observed whether it improved the performance of the system or not. The recognition system developed is composed of four blocks; the feature calculator, the phone-likelihood estimator, the speech decoder, and the natural language processor. The researchers investigated the performance of five different speech features through phoneme-level recognition. It was found out that among those tested, the Mel Frequency Cepstral Coefficients performed best, yielding a recognition rate of 73.57%. These speech features were then used in the feature calculation. Phone-likelihood estimation was done using a trained Multi-Layer Perception (MLP), and the output posterior probabilities were fed to a start-synchronous decoder which looked for the most probable word sequence. Tests in isolated word and continuous speech show that the system achieved a recognition rate of 71.12% and 9.95% respectively.For continuous-speech, NLP techniques on a syntactic level were employed to correct any errors in the output of the speech decoder. A stochastic word-level language model was used in this block. It verified if the generated word sequences are syntactically correct based on the Filipino language rules represented by the n-gram transitional probabilities. Finally, the proponents determined the viability of the implementing recognizer on a chip through system simulations in Simulink. Computational linguistics. Natural language processing (Computer science) Automatic speech recognition. Filipino speech recognition. Bayona, Michael Gringo Angelo R. Corrales, Patrick Simon T. FI UP UPD DENG-II LG 993.5 2009 E64 A46 Thesis