Development of a Femininity Estimator using Speaker Recognition Techniques for Voice Therapy of Gender Identity Disorder Clients

Abstract

This paper describes the development of an estimator of perceptual femininity (PF) of an input utterance using speaker recognition techniques. The estimator was designed for its clinical use and the target speakers are Gender Identity Disorder (GID) clients, especially MtF (Male to Female) transsexuals. The voice therapy for MtFs is composed of three kinds of training; 1) raising the baseline F 0 range, 2) changing the baseline voice quality, and 3) enhancing F 0 dynamics to produce an exaggerated intonation pattern. The first two focus on static acoustic properties of speech and the voice quality is mainly controlled by size and shape of the articulators, which can be acoustically characterized by the spectral envelope. Gaussian Mixture Models (GMM) of F 0 values and spectrums were built separately for biologically male speakers and female ones. Using the four models, PF was estimated automatically for each of 142 utterances of 111 MtFs. The estimated values were compared with the PF values obtained through listening tests. Results showed very high correlation (R=0.86), which is comparable to the intra-rater correlation.

Improved estimation of femininity using GMM supervectors and SVR for voice therapy of Gender Identity Disorder Clients

“…In our previous study [1], we built a male model and a female model separately for spectral features and pitch features. The four models were trained as GMMs using a large speech corpus containing biologically male speakers' utterances and biologically female speakers' ones.…”

“…For regression to perceptual femininity, simple linear regression of four scores was used in [1] (See Equation 3). When we use supervectors for regression, the number of parameters used in regression becomes remarkably large, which easily leads to the well-known overfitting problem.…”

“…In our previous study [1], two different kinds of GMMs were trained for each gender, which were a spectrum-based model and a pitch-based model. Similarly in this work, we build two UBM-GMMs, i.e., spectrum UBM-GMM and pitch UBM-GMM and they are adapted to each MtF speaker in the corpus.…”