Voice Conversion Based on Maximum-Likelihood Estimation of Spectral Parameter Trajectory

• Published in: IEEE transactions on audio, speech, and language processing Vol. 15; no. 8; pp. 2222 - 2235
• Main Authors: Toda, T., Black, A.W., Tokuda, K.
• Format: Journal Article
• Language: English
• Published: Piscataway, NJ IEEE 01.11.2007; Institute of Electrical and Electronics Engineers
• Subjects: Applied sciences; Conversion; Density; Dynamic feature; Exact sciences and technology; Frames; Gaussian mixture model (GMM); global variance; Information, signal and communications theory; Loudspeakers; Mathematical models; Maximum likelihood estimation; maximum-likelihood estimation (MLE); Mean square error methods; Natural languages; Parameter estimation; Signal processing; Spectra; Speech; Speech enhancement; Speech processing; Statistics; Telecommunications and information theory; Training data; Trajectories; Virtual colonoscopy; Voice; voice conversion (VC); Performance evaluation; Spectral method; Dynamic feature; maximum-likelihood estimation (MLE); Acoustic signal; voice conversion (VC); Mixture theory; Sound quality; Modeling; global variance; Mean square error; Degradation; Probability density; Accuracy; Gaussian process; Vocal signal; Gaussian mixture model (GMM); Maximum likelihood; Signal conversion
• ISSN: 1558-7916; 1558-7924
• DOI: 10.1109/TASL.2007.907344

In this paper, we describe a novel spectral conversion method for voice conversion (VC). A Gaussian mixture model (GMM) of the joint probability density of source and target features is employed for performing spectral conversion between speakers. The conventional method converts spectral parameters frame by frame based on the minimum mean square error. Although it is reasonably effective, the deterioration of speech quality is caused by some problems: 1) appropriate spectral movements are not always caused by the frame-based conversion process, and 2) the converted spectra are excessively smoothed by statistical modeling. In order to address those problems, we propose a conversion method based on the maximum-likelihood estimation of a spectral parameter trajectory. Not only static but also dynamic feature statistics are used for realizing the appropriate converted spectrum sequence. Moreover, the oversmoothing effect is alleviated by considering a global variance feature of the converted spectra. Experimental results indicate that the performance of VC can be dramatically improved by the proposed method in view of both speech quality and conversion accuracy for speaker individuality.