Discl-VC: Disentangled Discrete Representation and In-Context Learning for Controllable Zero-Shot Voice Conversion

0. Contents

Abstract
Demos -- zero-shot voice conversion with timbre transformation and prosody preservation
Demos -- zero-shot voice conversion with timbre and prosody transformation

1. Abstract

Currently, zero-shot voice conversion systems are capable of synthesizing the voice of unseen speakers. However, most existing approaches struggle to accurately replicate the speaking style of the source speaker or mimic the distinctive speaking style of the target speaker, thereby limiting the controllability of voice conversion. In this work, we propose Discl-VC, a novel voice conversion framework that disentangles content and prosody information from self-supervised speech representations and synthesizes the target speaker's voice through in-context learning with a flow matching transformer. To achieve fine-grained control over the generated speech's prosody, we introduce a mask generative transformer that predicts discrete prosody tokens in a non-autoregressive manner based on prompts. Experimental results demonstrate the superior performance of Discl-VC in zero-shot voice conversion and its remarkable accuracy in prosody control for synthesized speech.

Figure 1: The overall architecture of our proposed system.

Our system is trained in two stages:

In the first stage, we train all components except the Prosody Mask Transformer (PMT), and the training objective is: \(\mathcal{L}_{stage1} = \mathcal{L}_{Dur} + \mathcal{L}_{SimVQ} + \mathcal{L}_{FMT} + \mathcal{L}_{F0}\) .
In the second stage, we only train the PMT while keeping the other parts frozen, and the training objective is: \(\mathcal{L}_{stage2} = \mathcal{L}_{PMT}\) .