Effect of retroflex sounds on the recognition of Hindi voiced and unvoiced stops

摘要: As development of the speech recognition system entirely depends upon spoken language used for its development, and very fact that technology is highly dependent reverse engineering not possible, there an utmost need to develop such systems Indian languages. In this paper we present implementation a time delay neural network (TDNN) in modular fashion by exploiting hidden structure previously phonetic subcategory Hindi consonants. For study have selected all phonemes srecognition. A vocabulary 207 words was designed task-specific environment as database. phoneme, three-layered constructed trained using back propagation learning algorithm. Experiments were conducted categorize voiced, unvoiced stops, semi vowels, nasals fricatives. close observation confusion matrix stops revealed maximum retroflex with their non-retroflex counterparts.