Weight Zero Enhancement In Speech Synthesis Using Neural Networks
Gavin C. Cawley, Malcolm Iain Heywood, P.D. Noakes · 1993
In continuous speech the boundaries between allophones are not distinct, but are considerably blurred, an effect known as coarticulation, caused by the inertia of articulators (eg lips and tongue). Coarticulation carries little semantic content of an utterance, but we subconsciously expect to hear its effects in natural speech. This work is concerned with the training of neural networks for speech synthesis through generation of Linear Predictive Coding (LPC) parameters corresponding to the sequence of allophones. LPC attempts to find the coefficients A k of an all pole filter with a transfer function H(z), such that the spectral properties are optimally similar to that of a segment of sampled speech.