Improvement of endpoint detection for Thai isolated ord recognition

This paper proposes an approach for speech detection with mel-frequency cepstral coefficients (MFCC) for Thai isolated word recognition by increasing the difference between pre-frames and post-frames (Delta) to signal and distinguishing speech level in terms of speech energy, adaptive zero-crossing rate, and Euclidean distance of cepstrum-domain coefficients. Experimental result shows that the accuracy of speech recognition system gains the highest rate at 98.49% according to back-propagation neural network, from the database of 10 speakers by testing MFCC comparison of 3 different endpoint detection methods.