Cross gender voice morphing using Canonical Correlation Analysis

Abstract

Voice morphing one of the speech synthesis frameworks, in simplest term aim to transforms speaker's identity from source to target speaker while preserving the original content of message. This paper presents a novel spectral envelope mapping algorithm based on Canonical Correlation Analysis(CCA) that find the association between spectral envelope characteristics of source speaker and target speaker in terms of correlation as a similarity metric. Moreover, the speech also undergoes to prosodic modification using PSOLA as pitch frequency is also an important parameter for varying identity. This morphing algorithm is evaluated by taking the utterances from freely available CMU-ARCTIC speech dataset. The subjective experiment shows that the proposed method successfully transforms speaker identity and produced high-quality morphed signal.