Artificial Intelligent Systems and Machine Learning

Abstract

Automatic Speech Recognition (ASR) can be defined as an independent, computer-driven transcription of spoken language into readable text in real time. In this paper an ASR system was developed to recognize ten English isolated spoken digits. The work flow of the developed system starts with the speech signal being acquired from a microphone and brought into a Matlab development environment for analysis. Next a word-detection algorithm was used to separate each uttered digit from ambient noise and silence. Then a 13 Mel Frequency Cepstral Coefficients (MFCCs) were computed to represent the feature of each frame in the speech signal. Finally, an ensemble of 4 classifiers were combined in order to improve the overall performance of the classification. The ensemble consisted of Nearest Neighbor (1NN), five Nearest Neighbor (5NN), Dynamic Time Warping (DTW) and Minimum Euclidean Distance (MED). The outputs of these classifiers were combined by a majority voting method, to produce the final predicted class label. In the case of a tie .i.e. no majority, then a sensitivity measure (Se%) is calculated for each classifier. So if an unknown uttered digit was presented to the system, it would be assigned to the class predicted by the classifier which has the highest sensitivity value. The developed ASR system was trained and tested with a dataset of size 250 samples each. The ensemble based classification proved to be superior in terms of accuracy than any of the standalone classifiers.

Keywords

Full Text

References