Significance of frame size and frame shift on vowel on set point detection

In this paper, we analyse the effect of frame size and frame shift in detection of vowel on set point (VOP) under clean and noisy conditions, towards making VOP detection more accurate in practical scenario. For detection of VOP we use the state of art technique which combines the complementary evidences from excitation source, spectral peaks, and modulation spectrum. We carry out our experiments using TIMIT and NOISEX-92 database. In each case, 10 ms, 20 ms, and 30 ms frames with a frame shift of 25%, 50%, and 75% are adapted to evaluate the results. It is found that, 25% frame shift performs best irrespective of frame size. However, the frame size of 30 ms with a frame shift of 25% outperforms all the combinations used for detecting the VOPs. This combination shows an improvement of 2-4% approximately in total error rate (TER) over the baseline system, for both clean and noisy data.