| 摘要 |
<P>PROBLEM TO BE SOLVED: To solve the problem that, when learning an excitation model conventionally, since an error including periodicity generated during voiced filter characteristic estimation is mixed into a target signal used for estimating unvoiced filter characteristics and unvoiced filter characteristics are estimated excessively, an excitation signal is generated from the excessively estimated unvoiced filter characteristics while composing voices, and an excessive non-periodic component rides on a voice synthesized finally as noise. <P>SOLUTION: An unvoiced filter learning apparatus includes: an analysis residual signal receiving section for receiving one or more analysis residual signals; a non-periodic component signal extracting section for extracting one or more non-periodic component signals from one or more analysis residual signals; an unvoiced filter characteristic calculating section for calculating unvoiced filter characteristics on the basis of one or more non-periodic component signals; and an unvoiced filter characteristic storage section for storing unvoiced filter characteristics calculated by the unvoiced filter characteristic calculating section. Characteristics of an unvoiced filter can be appropriately learnt by the unvoiced filter learning apparatus. <P>COPYRIGHT: (C)2012,JPO&INPIT |
