DEVELOPMENT OF SPEECH RECOGNITION SYSTEM OF ISOLATED WORDS

Abstract

Speech recognition converts spoken words to text. This project objective is development of recognition systems that must be trained to a particular speech data set. Project uses a library of words/data set to train the system. The data set is composed of words from English language and different individuals. The basic technique we adopted is feature extraction from the speech signals of the words in the speech data set. Features are obtained using Linear Predictive Coding (LPC), Mel Frequency Cepstrum (MFC) & Statistical Functions. The complete feature set obtained is then used as an input to machine learning tool Weka, where the feature set is experimented against various machine learning algorithms. The best algorithm is opted keeping in view the results of the applied procedure in Weka. The best feature set is also separated from the complete feature set that has main differentiating parameters from the rest of the feature set. The best fit algorithm is applied on the offline input data used for training and experimentation for the best results. The feature set is stored for comparison and recognition. The online input signal is received through microphone used as the mainly input source. Its feature extraction is carried out on the guidelines of best feature set. The results are compared against the input database and maximum likelihood results are displayed.