Isolated Assamese Speech Recognition using Artificial Neural Network

This paper depicts a way of speech recognition of Assamese Isolated words in both speaker dependent and speaker independent cases by taking the parameters Zero Crossing Rate (ZCR), Short Time Energy (STE), and Mel Frequency Cepstral Coefficients (MFCC) that are extracted from the utterances of the Assamese words. The system builds with training phase, testing phase and the recognition phase. In the database, there consists of utterances of 100(hundred) frequently used Assamese isolated words whose syllable varies from 1 to 5 (monosyllabic to pentasyllabic) of 20(twenty) numbers of Assamese speakers with the same number of male and female ten speakers each, where each isolated word is spoken by twenty times by every speaker. The recognition/accuracy rate is high in each scenario. For speaker dependent speech recognition, we have got about 99% recognition rate, and in case of speaker independent speech recognition we have achieved 93 % recognition rate.