Some aspects of training methods in building a predictive classification model are studied in the context of discriminating between the stop consonants based on their places of articulation. Two training algorithms are taken-up: one from the likelihood training category and the other from the discriminative training category. These algorithms are discussed in a unified framework using information theoretic concepts to bring out their advantages and limitations with reference to the classification task. The experimental studies on recognition confirm the usefulness of exploiting these advantages for modelling the stop consonant transition region by a predictor. A waveform based scalar predictor, trained by the modified maximum likelihood (first) algorithm shows a substantial reduction in pitch pulse interference. The significance of discriminatively training the predictive classifier using a maximum mutual information type (second) algorithm is brought out in a spectral prediction framework.

Telugu characters can be visualised as being composed of circular segments of different radii. Recognition consists in segmenting the characters into the component elements and identifying them. We choose a feature set to preserve the canonical shapes while filtering out as noise the shape deviations encountered in real life. Hence, this approach does not require extensive training. Instead, "Feature Vector" parameters for individual "basic" characters are extracted from single specimens written in isolation. These are suitably combined to construct "Feature Vectors" for compound characters for the lexicon. These are compared with similar "Feature Vectors" extracted from the test samples to be recognised. Recognition scores ranged from 78 to 90% across different subjects (when the best match alone is taken) and from 91 to 95% for a single subject.

The approach described is based on an empirical parametric model for the handwriting recognition system. The parameters are so chosen and quantized as to retain only broad shape information, ignoring writer-dependent and other variability. Concatenation of character prototypes generates archetypal reference words for recognition, and training is unnecessary. The recognition scores exceed 90%.

We propose an efficient, self-organizing segmental measurement based on piecewise linear regression (PLR) fit of the short-term measurement trajectories. The advantages of this description are: (i) it serves to decouple temporal measurements from the recognition strategy; and, (ii) it leads to lesser computation as compared with conventional methods. Also, acoustic context can be easily integrated into this framework. The PLR measurements are cast into a stochastic segmental framework for phoneme classification. We show that this requires static classifiers for each regression component. Finally, we evaluate this approach on the phoneme recognition task. Using the TIMIT database. This shows that the PLR description leads to a computationally simple alternative to existing approaches.>

The k-nearest-neighbor decision rule is known to provide a useful nonparametric procedure for pattern classification. This rule is applied here to a vowel recognition problem and the effect of the number (k) of nearest neighbors, the size of the trained set and the type of the distance measure on vowel recognition performance is studied. It is shown that the vowel recognition performance remains approximately constant for all the values of k. The recognition performance initially improves with the size of the training set and then converges to an asymptotic value. Selection of a better distance measure leads to a significant improvement in vowel recognition performance.