Mitigating Hallucinations in LM-Based TTS Models via Distribution Alignment Using GFlowNets

Chenlin Liu, Minghui Fang, Patrick Zhang, Wei Zhou, Jie Gao, Jiqing Han

Introduction

Modern Text-to-Speech (TTS) systems have increasingly adopted the autoregressive "next-token prediction" paradigm characteristic of Large Language Models (LLMs). While this shift has enabled unprecedented levels of prosodic naturalness and speaker fidelity, it has also introduced a significant drawback: hallucinations. In speech synthesis, these manifest as skipped words, unnatural repetitions, or incoherent audio segments, particularly when processing complex or repetitive text.

A key contribution of this work is the empirical verification of the relationship between a model's internal state and its propensity to hallucinate. The researchers found a strong positive correlation between model uncertainty—measured by the entropy of the predicted token distribution—and the Character Error Rate (CER) of the generated speech.

Correlation between model uncertainty and log Character Error Rate (CER).

This finding suggests that hallucinations are not random failures but are preceded by detectable spikes in model uncertainty. By steering the model toward more deterministic, high-confidence generation paths, it is possible to preemptively mitigate these errors.

The GOAT Framework

To address these issues, the authors propose GOAT (GFlOwNet-guided distribution AlignmenT), a post-training framework designed to align the model's output distribution with high-reward (low-uncertainty) trajectories.

Overview of the GOAT framework, illustrating GFlowNet autoregressive sampling and the SubTB objective.

The framework treats speech generation as a flow optimization problem. Unlike standard Reinforcement Learning (RL) techniques that might cause the model to collapse into a single "greedy" output, GFlowNets learn to sample diverse trajectories proportional to their rewards. This preserves the natural variation inherent in human speech while filtering out erroneous outputs.

Key components of GOAT include:

  • Subtrajectory Balance (SubTB): An objective function that allows for fine-grained optimization of long sequences by breaking them into manageable segments.
  • Internal Reward Signal: Instead of requiring an external, computationally expensive reward model, GOAT utilizes the vanilla LM's own sharpened token probabilities as a reward signal.
  • Low-Rank Adaptation (LoRA): The optimization is applied via LoRA, ensuring that the improvements do not increase inference latency or require a complete model retraining.

Experimental Results

The GOAT framework was evaluated on the SeedTTS-Eval benchmark, with a specific focus on "test-hard" samples designed to trigger hallucinations (e.g., tongue twisters).

The results demonstrate significant improvements:

  • 50% Reduction in CER: On challenging test cases, GOAT halved the error rate compared to baseline models.
  • Stability Gains: Model uncertainty was reduced by up to 58%, leading to more consistent performance across different languages (English and Chinese).
  • Efficiency: Because the framework uses LoRA for post-training, the Real-Time Factor (RTF) remains unchanged, making it suitable for production environments.

Conclusion

GOAT provides a robust, efficient method for mitigating hallucinations in LM-based TTS systems. By leveraging the detectable link between uncertainty and error, and using GFlowNets for distribution alignment, the framework achieves substantial reliability gains without sacrificing the speed or naturalness of modern autoregressive speech synthesis.

Summary updated

BibTeX
@inproceedings{liu2025mitigating,
  author = {Chenlin Liu and Minghui Fang and Patrick Zhang and Wei Zhou and Jie Gao and Jiqing Han},
  title = {{Mitigating Hallucinations in LM-Based TTS Models via Distribution Alignment Using GFlowNets}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure