Runtime and speech quality survey of a voice conversion method

摘要

Several methods for voice conversion have been established. The research aims at the characteristics of a target speaker and a near-to-natural speech quality. This contribution summarizes the listening experiments with four conversion methods including the assessment of speech quality, listening effort and similarity to the target voice. The subjective evaluation of similarity is checked by an instrumental distance measure based on logarithmic spectral distortion. Practical applications of voice conversion require an appropriate runtime performance and memory use. We select a conversion method based on VTLN to demonstrate the runtime and quality trade-off. In the case example, we survey the quality assessment depending on different training constellations with a varied data amount and training time. Furthermore, we discuss the runtime performance of the selected conversion method under typical operating conditions. The experiments cover the influence of system resources, setting of conversion parameters (warping factors) and different training constellations. The observed real-time factors of a non-optimized laboratory VC version are inappropriate for typical application scenarios.