Consonant-Vowel (CV) segments are the basic production units in Indian languages. It was found that about 45% of CVs that occur in a text are Stop-Consonant-Vowel (SCV) segments. Recognition of frequently occurring SCVs is important for developing a vocabulary independent continuous speech recognition system for Indian languages. In this paper, we discuss some issues in recognition of SCVs in continuous speech of Hindi. Suitable parametric representation and classifier models are needed in order to discriminate a large number of confusable SCV classes in Indian languages. We compare the recognition performance of a subset of SCVs for three classifier models, namely, multilayer perceptron, time delay neural network and hidden Markov model. Classification performance of about 70% for multispeaker data and about 60% for speaker independent data has been obtained for ten confusable SCV classes. Analysis of the recognition performance indicates that many of the recognition errors are due to incorrect classification of the Place of Articulation (POA) of the stop consonants in SCVs. The clues for POA are present in the transition region of an SCV segment and hence the representation of these regions becomes crucial for recognition of SCVs. We compare different parametric representations for classification of the transitions in SCVs using multilayer perceptron as the classifier model. © 1996 Taylor & Francis Group, LLC.