Comparison of vowel recognition systems based on different extracted features.

摘要

The thesis revolves around development and comparison of vowel recognition systems. For automatic speech recognition systems, a preprocessing step called the endpoint detection is required to isolate the speech of interest from the background noise so as to be able to create a speech pattern or template.; Particular features are extracted from the speech templates. The features extracted in this thesis are: (1) Energy at different frequency bands attained by Wavelet packet decomposition. (2) Formant frequencies of vowel utterances.; Approaches like frequency partitioning and LPC models are developed to extract features. The two features are compared by feeding them into neural network classifiers. The neural networks are developed by adopting the concepts of vector quantization and K-means algorithms. The recognition systems built on the basis of neural networks, are tested and compared for their relative efficiencies.; The algorithms are implemented, simulated and tested in MATLAB.