The Codec Language Model-based Zero-Shot Spontaneous Style TTS System for CoVoC Challenge 2024

TL;DR

This paper presents a zero-shot spontaneous style TTS system using a LLaMA-based codec language model with guidance strategies, achieving top naturalness and quality in the CoVoC 2024 challenge.

Contribution

It introduces a novel LLaMA-based codec language model with delay pattern and classifier-free guidance for spontaneous style voice cloning in zero-shot TTS.

Findings

Achieved a MOS of 3.80 for speech naturalness

Outperformed in speech quality and speaker similarity

Effective data preprocessing improved utterance quality

Abstract

This paper describes the zero-shot spontaneous style TTS system for the ISCSLP 2024 Conversational Voice Clone Challenge (CoVoC). We propose a LLaMA-based codec language model with a delay pattern to achieve spontaneous style voice cloning. To improve speech intelligibility, we introduce the Classifier-Free Guidance (CFG) strategy in the language model to strengthen conditional guidance on token prediction. To generate high-quality utterances, we adopt effective data preprocessing operations and fine-tune our model with selected high-quality spontaneous speech data. The official evaluations in the CoVoC constrained track show that our system achieves the best speech naturalness MOS of 3.80 and obtains considerable speech quality and speaker similarity results.