A Wavelet-Based Method for Morphing Audio Recordings of Interjections from One Voice to Another

This article is dedicated to a new method for morphing audio recordings of interjections: the original speaker's voice is transformed to the target one within a given voice pair, with preserving the original intonation. In speech synthesis solutions for Russian, interjections are poorly covered: the intonation set is limited, the contrary to human voice richness. Introducing a variety of interjections for communicative robots and voice assistants is of interest, as this would enable them to appear more natural in human-machine interaction, as well as to overcome existing constraints of dialogue systems and voice interfaces which lack affection. We designed an interjections synthesizing method aimed at fulfilling the F-2 communicative robot’s reactions database. For verifying the applicability of synthesized recordings, we matched interjections-based reactions to input interjections, in the manner of Campbell’s classic examples of dialogues limited to interjections, typical in colloquial human speech. We invited respondents to interact with F-2, to give cues and subsequently estimate answers. Collected evaluation data are discussed; most synthesized interjections are perceived as natural. Thus our research procedure enabled us to enrich the variety of our robot’s communication means, and to check the appropriateness of a number of reacting scenarios: several reactions are dropped out as inappropriate, according to our respondents’ opinions. The introduced wavelet-based method is applicable for morphing interjections from various speakers to the target voice, including the cases of multimodal corpora utilized for developing human-like reactions for affective communication solutions.