First International Conference on Spoken Language Processing (ICSLP 90)
We trained a three-layered neural network to discriminate between the left and right contexts around the burst point. The shifting window is the portion of the input speech which will serve as input to the network at a time. The detection window is the portion of the input speech where the searched burst locates. The outputs from our network distinguish three states, i. e. before or after the burst and outside the burst of the shifting window. The acoustic properties used was speech power time series of a 5-band me 1-sea 1ed LPC spectrum. The network consisted of 80 input units 30 hidden units and 3 output units. French voiced stop consonants /b,d,g/ served as input. Elimination of phantom burst point was attained. The decisions error was around 15~ms. Usually right decisions consistently proceeded the real burst point and then followed by left decisions. High frequency component was effective among 5 frequency bands.
Bibliographic reference. Kitazawa, Shigeyoshi / Serizawa, Masahiro (1990): "An artificial neural network for the burst point detection", In ICSLP-1990, 1069-1072.