Training Robust Zero-Shot Voice Conversion Models with Self-supervised Features

TL;DR

This paper introduces a robust zero-shot voice conversion method leveraging self-supervised speech features, a length resampling decoder, and additional training techniques to improve speaker similarity and content preservation without parallel data.

Contribution

It proposes a novel length resampling decoder and training strategies that enhance zero-shot voice conversion quality using self-supervised features without changing the core architecture.

Findings

Outperforms baselines on VCTK dataset

Achieves high speaker similarity and content preservation

Best performance on LibriTTS with proposed techniques

Abstract

Unsupervised Zero-Shot Voice Conversion (VC) aims to modify the speaker characteristic of an utterance to match an unseen target speaker without relying on parallel training data. Recently, self-supervised learning of speech representation has been shown to produce useful linguistic units without using transcripts, which can be directly passed to a VC model. In this paper, we showed that high-quality audio samples can be achieved by using a length resampling decoder, which enables the VC model to work in conjunction with different linguistic feature extractors and vocoders without requiring them to operate on the same sequence length. We showed that our method can outperform many baselines on the VCTK dataset. Without modifying the architecture, we further demonstrated that a) using pairs of different audio segments from the same speaker, b) adding a cycle consistency loss, and c) adding a speaker classification loss can help to learn a better speaker embedding. Our model trained on LibriTTS using these techniques achieves the best performance, producing audio samples transferred well to the target speaker's voice, while preserving the linguistic content that is comparable with actual human utterances in terms of Character Error Rate.