sprocket: Open-Source Voice Conversion Software

Abstract

Statistical voice conversion (VC) is a technique to convert specific non-or paralinguistic information while keeping linguistic information unchanged, and speaker conversion has been studied as a typical application of VC for a few decades. To better understand various VC techniques using a freely available common dataset, the Voice Conversion Challenge (VCC) was launched in 2016 and the 2nd challenge was held in 2018. As one of the baseline systems for VCC 2018, we developed open-source VC software called "sprocket", in which not only conventional techniques, such as a trajectory-based conversion method using a Gaussian mixture model (GMM) and a vocoderbased conversion framework but also recently developed techniques, such as a vocoder-free VC framework, have been implemented. Using sprocket, it is possible to 1) easily reproduce converted voices using the VCC datasets and 2) develop VC systems using other parallel speech datasets with fundamental VC functions, such as acoustic feature extraction, time alignment between the source and target features, GMM training, feature conversion, and waveform generation. In this paper, we describe 1) the technical details and use of sprocket, 2) the development of the baseline systems for the HUB and SPOKE tasks of VCC 2018 using sprocket, and 3) the performance of sprocket as a VC system by demonstrating results for our developed baseline systems in VCC 2018.

NU Voice Conversion System for the Voice Conversion Challenge 2018

“…The results of using mel-cepstral distortion to evaluate the spectral conversion module are given in the objective evaluation results. An internal subjective evaluation was conducted to assess the performance of the NU VC system with the provided baseline system, i.e., "sprocket" [27], where the results are given in the internal subjective evaluation section. Finally, the last three sections describe the official results of the subjective evaluation in VCC 2018.…”

Noise-Robust Voice Conversion Using High-Quefrency Boosting via Sub-Band Cepstrum Conversion and Fusion

“…The GMM is trained using joint vectors of X i and Y i in the parallel data set, which have been automatically aligned to each other by dynamic time warping (DTW) [4]. The detailed steps can be found in References [2,25]. DIFFGMM is a differential Gaussian mixture model.…”

“…For a more detailed conversion process of VC and DIFFGMM and the training process of parallel VC based on GMM, please refer to Figures 1 and 2 in Kobayashi and Toda [25].…”

NU Voice Conversion System for the Voice Conversion Challenge 2018

“…An internal subjective evaluation was conducted to assess the performance of the NU VC system with the provided baseline system, i.e., "sprocket" [27], where the results are given in the internal subjective evaluation section. Finally, the last three sections describe the official results of the subjective evaluation in VCC 2018.…”

“…In the internal subjective evaluation, two preference tests (naturalness and speaker similarity) were conducted to compare the performance of the NU VC system with that of the baseline system, i.e., sprocket [27]. All 16 speaker pair models for the four source and four target speakers were used in the evaluation.…”