Thai speech recognition using double filter banks for basic voice commanding

Abstract

This paper describes the methodology to recognize Thai speech words by integrating two approaches e.g., Double filter banks and Euclidian distance in a feature extraction and the recognition processes, respectively. Firstly, the speech signals are transformed into the 3-dimension of signal or spectrogram. The spectrogram displays energy information along both time and frequency axes. Secondly, the frequencies to be within the bin spread and correlated them with each triangular filter. Thus, each bin holds a weighted sum, and represents the spectral magnitude in that filter bank channel. Finally, the filter banks are normalized into the normalized bank for comparison between the entry signal of word and various words of dictionary. The Euclidian distance is used to measure the similarity between them. The system was evaluated for its accuracy and stability in performing various conditions. The accuracy was tested with 9, 000 speeches from several volunteers. The average accuracy rate is about 96.3 %. The results show that the evaluation was beyond satisfaction for every aspect. © 2010 IEEE.