Creating synthetic voices for children by adapting adult average voice using stacked transformations and VTLN

Abstract

This paper describes experiments in creating personalised children's voices for HMM-based synthesis by adapting either an adult or child average voice. The adult average voice is trained from a large adult speech database, whereas the child average voice is trained using a small database of children's speech. Here we present the idea to use stacked transformations for creating synthetic child voices, where the child average voice is first created from the adult average voice through speaker adaptation using all the pooled speech data from multiple children and then adding child specific speaker adaptation on top of it. VTLN is applied to speech synthesis to see whether it helps the speaker adaptation when only a small amount of adaptation data is available. The listening test results show that the stacked transformations significantly improve speaker adaptation for small amounts of data, but the additional benefit provided by VTLN is not yet clear.

A Situational Analysis of Current Speech-Synthesis Systems for Child Voices: A Scoping Review of Qualitative and Quantitative Evidence

“…However, it was refreshing to discover studies focusing on other variants of English such as Irish English [41] and Indian English [19,42]. Although less common, researchers also considered other languages when experimenting with child-speech synthesis, including Norwegian [5], Spanish [18], Punjabi [43], Finnish [44], German [45,46], Czech and Slovak [47], Mandarin [27,31] and quite often, Italian [21,29,30,[48][49][50][51][52].…”

“…Interestingly, aside from creating a child voice by adapting an average adult voice or an average child voice, Karhila et al's study [44] compared two additional adaptation methods using stacked transformations: StA and StVA. In the first method, StA, an average voice trained from adult data was adapted using training data of the average child voice.…”

“…This model was then further adapted to resemble a specific target child speaker. In the second method, StVA, an average voice trained from adult data was adapted using training data of the average child voice, and then VTLN occurred [44]. It was found that stacked transformation systems (StA and StVA) were preferred by listeners and resulted in better adapted voices than directly adapting the average adult voice or the child voice [44].…”

“…In addition, many researchers have attempted to build child-speech models by adapting adult-speech models. This has been proven to be a viable method when there are limited speech data available [5,18,[23][24][25][26]37,44,48]. In a study by Hagen, Pellom and Hacioglu [18], a synthetic children's model was derived without child-speech data by using adult-speech data.…”

Voice Conversion for Dubbing Using Linear Predictive Coding and Hidden Markov Model

“…Another technique is adaptive voice conversion, which can be used to dub children's voices in children's movie applications. Techniques for creating or generating children's voices have been proposed in the study [12], [13], [14], [15]. Watts et al [14], [15] proposed the Hidden Markov Model (HMM) as a basis for a method of synthesizing children's voices.…”