HNM parameter transform for voice conversion using a HMM-WDLT framework

This paper presents a Harmonic + Noise Model (HNM)-based voice conversion technique under a Hidden Markov Model-Weighted Deviation Linear Transformation (HMM-WDLT) framework. In a comparative study of four methods of converting the extracted line spectral frequency (LSF) parameters of one speaker to another, the HMM-WDLT achieves the lowest average spectral distortion. A remedial process is developed to enhance the formant characteristics while preserving the variance of the LSF parameters. The frame duration, manifested by the slope of the dynamic time warping (DTW) path, is regarded as an output variable of the conversion function. To take full advantage of the attributes of the HNM, the conversion algorithm fine-tunes the harmonic magnitudes below 2 kHz for each critical band. Listening test reveal that the converted speech successfully catches the speaker´s individuality with satisfactory quality.