Hierarchical Decoding for Discrete Speech Synthesis with Multi-Resolution Spoof Detection

TL;DR

This paper introduces MSpoof-TTS, a training-free hierarchical decoding framework that uses multi-resolution spoof detection to improve zero-shot discrete speech synthesis quality and robustness.

Contribution

It proposes a novel multi-resolution spoof detection method integrated into hierarchical decoding, enhancing speech synthesis without retraining or preference optimization.

Findings

Improved perceptual realism in zero-shot speech synthesis.

Effective detection of unnatural patterns at multiple temporal scales.

Enhanced robustness against token-level artifacts.

Abstract

Neural codec language models enable high-quality discrete speech synthesis, yet their inference remains vulnerable to token-level artifacts and distributional drift that degrade perceptual realism. Rather than relying on preference optimization or retraining, we propose MSpoof-TTS, a training-free inference framework that improves zero-shot synthesis through multi-resolution spoof guidance. We introduce a Multi-Resolution Token-based Spoof Detection framework that evaluates codec sequences at different temporal granularities to detect locally inconsistent or unnatural patterns. We then integrate the spoof detectors into a hierarchical decoding strategy, progressively pruning low-quality candidates and re-ranking hypotheses. This discriminator-guided generation enhances robustness without modifying model parameters. Experiments validate the effectiveness of our framework for robust and high-quality codec-based speech generation. Audio samples are available at https://danny-nus.github.io/MSpoofTTS.github.io/.