LSP speech synthesis using backpropagation networks

A multilayer perceptron (MLP) similar to that used in the NETtalk system is used to form a mapping between sequences of allophones and corresponding frames of LPC synthesizer control parameters. Three parameter sets equivalent to the LPC coefficients, line spectral pair (LSP), PARCOR and log area ratio, are evaluated. In addition to training a standard MLP, networks which have been decomposed according to phonetic class and by allophone, are trained. Decomposition is found to reduce training time and produce greater accuracy on the training set, however the network decomposed by allophone is found to receive too few training patterns to generalize properly on new data.<
>