| 摘要 |
<p>The method is creating a speech output from a succession of input linguistic target elements including target characteristics, where the speech output is formed by concatenating a sequence of selected waveform units, each selected waveform unit corresponding to an input linguistic target element. The method includes repeating iterative sequences of forward steps, backward steps and the creating of speech output until the forward steps have reached the final target element. The same optimal sequence of selected waveform units for all target elements of a succession of input linguistic target elements starting with an initial target element and ending with a final target element as the standard Viterbi search are emitted but the optimal units become available in a pipelined manner without requiring the calculation of path costs for the final target element and without complete backtracking form the final to the initial target element. The latency, i.e. the amount of computation time before outputting selected waveform units for a beginning part of the target sequence is much shorter than in a Viterbi search.</p> |
