Augmented speech production based on real-time statistical voice conversion

Abstract

In human-to-human speech communication, various barriers are caused by some constraints, such as physical constraints causing vocal disorders and environmental constraints making it hard to produce intelligible speech. These barriers would be overcome if our speech production was augmented so that we could produce speech sounds as we want beyond these constraints. Voice conversion (VC) is a technique for modifying speech acoustics, converting non-/para-linguistic information to any form we want while preserving the linguistic content. One of the most popular approaches to VC is based on statistical processing, which is capable of extracting a complex conversion function in a data-driven manner. Although this technique was originally studied in the context of speaker conversion, which converts the voice of a certain speaker to sound like that of another specific speaker, it has great potential to achieve various applications beyond speaker conversion. This paper briefly reviews a trajectory-based conversion method that is capable of effectively reproducing natural speech parameter trajectories utterance by utterance and highlights several techniques that extend this trajectory-based conversion method to achieve real-time conversion processing. Finally this paper shows some examples of real-time VC applications to enhance human-to-human speech communication, such as speaking-aid, silent speech communication, and voice changer/vocal effector.