Age Speech Conversion Based On Spectrum And Prosodic Features

Age speech conversion means that source age speech spoken by one speaker sounds like the target age speech spoken by another speaker through changing age information of the speech while keeping the speech content unchanged. Age speech conversion is speaker-independent, that means the conversion model is suitable for all speakers of the same age period. It is different from common speech conversion.To research acoustic feature parameters which related to age information well, a comprehensive age speech corpus is established based on four age periods which are children, youth, middle age and old age period. On this basis, a new method using UBM groups of short-time spectrum and prosodic features is also proposed in speaker-independent age speech conversion. To reduce the differences of speech spoken by different speakers in the same age period, the paper makes GMM models by spectrum parameters of different speakers in the same age period and UBM groups are obtained by self-adapting cluster. A set of spectrum conversion functions is derived by joint training for the spectrum of Each UBM in UBM groups and target age period. Then the likelihood of test voice in each UBM of UBM groups in conversion phase is calculated and the optimal spectrum conversion function is got by maximum likelihood criterion. After that, formants adjustment is conducted for decreasing the losing and obscure of important spectrum information. On the other hand, prosodic features which have great influence for age speech conversion are selected for prosodic feature conversion. Finally, STRAIGHT is used to synthesize converted age speech.The results of experiments based on objective and subjective evaluation are showed that the method proposed in this paper makes the converted speech more inclined to the target age period while keeping the voice quality. The age speech conversion has a good universality and doesn’t need train repeatedly, so the system is effective and flexible.