Experimental investigation on the efficacy of Affine-DTW in the quality of voice conversion

In this paper, the performance of Affine-DTW, which performs appropriate time alignment between source and target features in voice conversion (VC), is experimentally and thoroughly investigated. In traditional VC, parallel data are often required to train a mapping model between source and target features. While VC with non-parallel data is also studied to avoid collecting parallel data, the quality of its converted speech is still inferior to the traditional one with parallel data. One approach to further progress in VC is exploiting both parallel and non-parallel data, the former of which is pre-stored and the latter of which is assumed to be easily collected. In this case, it is still worthwhile to study time-alignment techniques to obtain appropriate alignment of parallel data. Affine-DTW is a technique in which dynamic time warping (DTW) and coarse conversion based on affine transformation are iteratively performed. In Affine-DTW, time alignment and parameters of affine transformation can be analytically calculated so that it can be easily adopted as pre-processing in VC. However, the influence on the performance of trained models based on the obtained alignments has not been well investigated experimentally. Hence, this paper investigates the performance of Affine-DTW in terms of quality improvement of converted speech in traditional VC methods based on Gaussian mixture models, non-negative matrix factorization and neural networks. Experimental results show that Affine-DTW obtains appropriate alignments and the naturalness improvement of converted speech in subjective assessments is observed in trained models based on the alignments.