Sure, here's a blog-style summary of the paper "Mitigating Hallucinations in LM-Based TTS Models via Distribution Alignment Using GFlowNets":
TL;DR
This paper introduces GOAT, a novel post-training framework that uses GFlowNets to reduce "hallucinations" (unwanted deviations from the input text) in LM-based Text-to-Speech (TTS) models. By framing TTS generation as a distribution alignment problem and leveraging uncertainty analysis, GOAT significantly improves speech quality without requiring additional training data or increasing inference latency.
What Problem Does This Paper Address?
Large Language Model (LLM)-based Text-to-Speech (TTS) systems have achieved impressive results in generating natural-sounding speech. However, they suffer from a phenomenon called "hallucination," where the generated speech deviates from the input text. This can manifest as mispronunciations, missing words, or even entirely fabricated content. Existing methods to combat this either demand excessive computational resources during training or introduce significant latency during inference, making them impractical for real-time applications. This paper aims to address this by proposing a post-training method that mitigates hallucinations efficiently.
Methods
The core of GOAT lies in its novel approach to optimizing the TTS generation process. The authors first conduct an uncertainty analysis, revealing a strong correlation between model uncertainty and hallucination. They then reframe TTS generation as a trajectory flow optimization problem, leveraging Generative Flow Networks (GFlowNets).
Key techniques include:
- Distribution Alignment: GOAT aims to align the model's output distribution with a target distribution that favors high-confidence sequences, thereby reducing uncertainty.
- GFlowNets: This probabilistic framework is used to learn an optimal transition process from an initial state to a terminal state, effectively guiding the generation towards desired outcomes.
- Enhanced Subtrajectory Balance (SubTB) Loss: This loss function allows the model to learn from subsequences of varying lengths, enabling fine-grained optimization.
- Intrinsic Reward Function: GOAT designs an internal reward function based on the LM's token sampling probability, eliminating the need for external reward models and enabling training on unlabeled data.
- Reward Temperature Decay and Learning Rate Optimization: These strategies are employed to balance performance and training stability, suppressing reward hacking (where the model exploits the reward function for unintended gains).
Datasets
- Training: The authors used the
train-clean-100subset of the LibriTTS corpus for English and the premium subset of WenetSpeech4TTS for Chinese. They also created a mixed dataset by combining 500 samples from each. - Evaluation: The official
seed-tts-evalbenchmark was used, which includes thetest-en(Common Voice),test-zh(DiDiSpeech-2), and a challengingtest-hardsubset.
Key Insights and Ideas
- Uncertainty as a Hallucination Proxy: The paper provides strong empirical evidence that model uncertainty is a reliable indicator of potential hallucinations in LM-based TTS. This insight is crucial for developing targeted mitigation strategies.
- GFlowNets for TTS Optimization: Applying GFlowNets to the sequence-level generation of TTS models offers a novel perspective for improving output quality and reducing errors.
- Post-Training Mitigation is Viable: GOAT demonstrates that effective hallucination mitigation can be achieved after the initial model training, without the need for extensive retraining or massive datasets.
- Balancing Performance and Stability: The careful design of reward temperature decay and learning rate optimization highlights the importance of these techniques in preventing reward hacking and ensuring robust training.
Key Results
- GOAT reduces character error rates (CER) by over 50% on challenging test cases.
- It lowers model uncertainty by up to 58%.
- The method achieves comparable or superior performance to the baseline on content consistency (CER/WER), speaker similarity (SS), and speech quality (UTMOS) metrics.
- GOAT significantly outperforms low-temperature sampling strategies for hallucination suppression, surpassing them by over 30% on CER/WER.
- The framework demonstrates strong robustness and generalization capabilities across different languages and datasets.
- Inference latency is not significantly increased compared to the baseline.
Limitations and Future Work
- Nuance of Hallucinations: The authors acknowledge that hallucination phenomena are more nuanced than uncertainty alone can capture, and GOAT currently addresses only a subset of cases.
- Model Specificity: The evaluation has primarily focused on Cosy Voice 2. Future work will extend GOAT to cover a broader range of hallucinations and assess its effectiveness across diverse LM-based TTS architectures.
- Broader Applications: The paper suggests that future work could explore more efficient training strategies and investigate the effectiveness of GOAT even with weaker or no prompts.
Summary updated