Isolated Word Recognition Based on Power Normalized Cepstrum and Machine Learning Clusters

Speech recognition is the ability of a machine to convert the words spoken by humans to text. Speech recognition is necessary for communication between the human and machine. This project is about recognizing isolated words such as digits from one to nine, enter, erase, no, go, start, stop, yes, etc. Speech recognition is done in two phases: Training and Testing. This work is done in several steps, such as collection of data, feature extraction, clustering of feature vectors and storing the models. Decision-making is done by pattern matching. Collected data is divided into train data and test data for each word. During the training phase, the feature extraction is performed using the Power Normalized Cepstral Coefficients (PNCC) algorithm and grouping all these features by a clustering algorithm. For the testing phase, the PNCC features are extracted. The distance between the training data clusters and testing data features is determined for each word. The minimum length is determined for each test vector, and an average of all the minimum distances is calculated. Recognition accuracy is a performance metric, and the system’s overall accuracy is 90%.