Selected work

Representative Papers

RobustSpeechFlow: Learning Robust Text-to-Speech Trajectories via Augmentation-based Contrastive Flow Matching

May 21, 2026

This work addresses the issue of word skipping and repetition in flow-matching-based text-to-speech (TTS) synthesis, which arises from inaccurate alignments. The authors propose a latent-space augmentation strategy that explicitly models failure modes without requiring external aligners or preference data, while preserving the original input length. This approach is integrated into a contrastive flow-matching framework and represents the first application of augmentation-based contrastive flow matching to enhance TTS content fidelity. It seamlessly fits into existing zero-shot TTS pipelines. Experimental results demonstrate consistent improvements: on Seed-TTS-eval, the word error rate (WER) decreases from 1.44% to 1.38%; on the ZERO500 benchmark, character error rates (CER) for English and Korean drop from 0.48% and 0.81% to 0.35% and 0.57%, respectively, with 24 function evaluations (NFE).

0 citationsRead paper

Robust TTS Training via Self-Purifying Flow Matching for the WildSpoof 2026 TTS Track

Dec 19, 2025

To address label noise and environmental diversity in real-world speech scenarios within the WildSpoof 2026 Text-to-Speech (TTS) Track, this paper proposes the Self-Purifying Flow Matching (SPFM) framework. SPFM introduces, for the first time in TTS flow matching, an explicit noisy sample routing mechanism that dynamically identifies and isolates suspicious text–speech pairs while preserving their acoustic information for unconditional flow matching training. Integrated with the open-source Supertonic model, SPFM jointly optimizes conditional and unconditional flow matching losses, employs lightweight fine-tuning, and adopts dynamic sample weighting. Experiments demonstrate that SPFM achieves the lowest Word Error Rate (WER) on the WildSpoof TTS Track, while attaining second-best perceptual scores on both UTMOS and DNSMOS—substantially improving robustness to label noise and generalization across diverse acoustic conditions.

0 citationsRead paper

Training Flow Matching Models with Reliable Labels via Self-Purification

Sep 23, 2025

Label noise—such as human annotation errors—severely degrades model performance. To address this, we propose the first self-purifying training framework tailored for Flow Matching (FM), requiring neither pretraining nor auxiliary modules. During training, it dynamically estimates label credibility per sample and filters unreliable instances. Our core innovation lies in embedding the self-purification mechanism directly into FM’s continuous-time modeling: it enables noise-aware gradient updates and adaptive learning-weight adjustment based on estimated label reliability. Experiments on the real-world noisy speech dataset TITW demonstrate that our method significantly improves the accuracy and fidelity of generated samples with respect to conditioning information. It outperforms existing baselines in robustness against label noise, establishing a new state-of-the-art for noise-robust flow-based generative modeling.

0 citationsRead paper

Improving Test-Time Performance of RVQ-based Neural Codecs

Sep 23, 2025

RVQ-based neural audio codecs suffer from high quantization error and limited synthesis quality during inference due to fixed codebooks. Method: This paper proposes a test-time dynamic code selection algorithm that requires no retraining. At inference, it jointly optimizes discrete codebook indices across all RVQ levels via hierarchical codebook search and layer-wise quantization error minimization, overcoming the suboptimality of conventional greedy encoding. Contribution/Results: The core innovation lies in formulating codebook selection as a differentiable path optimization problem—preserving discrete constraints while enabling end-to-end gradient backpropagation of reconstruction error. Experiments demonstrate significant improvements in both objective metrics (e.g., LSD, MRSTFT) and subjective MOS scores, with an average 23.6% reduction in quantization error. The method is fully compatible with existing RVQ models and incurs negligible deployment overhead.

0 citationsRead paper

Length-Aware Rotary Position Embedding for Text-Speech Alignment

Sep 14, 2025

Existing Transformer-based text-to-speech (TTS) systems commonly employ Rotary Position Embedding (RoPE), which relies on absolute token indices and thus struggles to model the dynamic alignment between text and speech—particularly under variable speech durations or long-sequence generation, leading to significant performance degradation. To address this, we propose Length-Aware Rotary Position Encoding (LARoPE), which explicitly encodes relative query-key distances via length-normalized positional indices, thereby enhancing robustness in cross-modal alignment. Integrated into the Transformer’s cross-attention mechanism, LARoPE requires no additional parameters while improving stability in long-sequence modeling. Experiments demonstrate that LARoPE accelerates convergence, improves alignment accuracy and speech naturalness, achieves state-of-the-art word error rate on zero-shot TTS benchmarks, and enables high-fidelity synthesis of utterances up to 30 seconds in duration.

0 citationsRead paper

Recent publications

Latest Papers

RobustSpeechFlow: Learning Robust Text-to-Speech Trajectories via Augmentation-based Contrastive Flow Matching

May 21, 2026

This work addresses the issue of word skipping and repetition in flow-matching-based text-to-speech (TTS) synthesis, which arises from inaccurate alignments. The authors propose a latent-space augmentation strategy that explicitly models failure modes without requiring external aligners or preference data, while preserving the original input length. This approach is integrated into a contrastive flow-matching framework and represents the first application of augmentation-based contrastive flow matching to enhance TTS content fidelity. It seamlessly fits into existing zero-shot TTS pipelines. Experimental results demonstrate consistent improvements: on Seed-TTS-eval, the word error rate (WER) decreases from 1.44% to 1.38%; on the ZERO500 benchmark, character error rates (CER) for English and Korean drop from 0.48% and 0.81% to 0.35% and 0.57%, respectively, with 24 function evaluations (NFE).

0 citationsRead paper

Robust TTS Training via Self-Purifying Flow Matching for the WildSpoof 2026 TTS Track

Dec 19, 2025

To address label noise and environmental diversity in real-world speech scenarios within the WildSpoof 2026 Text-to-Speech (TTS) Track, this paper proposes the Self-Purifying Flow Matching (SPFM) framework. SPFM introduces, for the first time in TTS flow matching, an explicit noisy sample routing mechanism that dynamically identifies and isolates suspicious text–speech pairs while preserving their acoustic information for unconditional flow matching training. Integrated with the open-source Supertonic model, SPFM jointly optimizes conditional and unconditional flow matching losses, employs lightweight fine-tuning, and adopts dynamic sample weighting. Experiments demonstrate that SPFM achieves the lowest Word Error Rate (WER) on the WildSpoof TTS Track, while attaining second-best perceptual scores on both UTMOS and DNSMOS—substantially improving robustness to label noise and generalization across diverse acoustic conditions.

0 citationsRead paper

Training Flow Matching Models with Reliable Labels via Self-Purification

Sep 23, 2025

Label noise—such as human annotation errors—severely degrades model performance. To address this, we propose the first self-purifying training framework tailored for Flow Matching (FM), requiring neither pretraining nor auxiliary modules. During training, it dynamically estimates label credibility per sample and filters unreliable instances. Our core innovation lies in embedding the self-purification mechanism directly into FM’s continuous-time modeling: it enables noise-aware gradient updates and adaptive learning-weight adjustment based on estimated label reliability. Experiments on the real-world noisy speech dataset TITW demonstrate that our method significantly improves the accuracy and fidelity of generated samples with respect to conditioning information. It outperforms existing baselines in robustness against label noise, establishing a new state-of-the-art for noise-robust flow-based generative modeling.

0 citationsRead paper

Improving Test-Time Performance of RVQ-based Neural Codecs

Sep 23, 2025

RVQ-based neural audio codecs suffer from high quantization error and limited synthesis quality during inference due to fixed codebooks. Method: This paper proposes a test-time dynamic code selection algorithm that requires no retraining. At inference, it jointly optimizes discrete codebook indices across all RVQ levels via hierarchical codebook search and layer-wise quantization error minimization, overcoming the suboptimality of conventional greedy encoding. Contribution/Results: The core innovation lies in formulating codebook selection as a differentiable path optimization problem—preserving discrete constraints while enabling end-to-end gradient backpropagation of reconstruction error. Experiments demonstrate significant improvements in both objective metrics (e.g., LSD, MRSTFT) and subjective MOS scores, with an average 23.6% reduction in quantization error. The method is fully compatible with existing RVQ models and incurs negligible deployment overhead.

0 citationsRead paper

Length-Aware Rotary Position Embedding for Text-Speech Alignment

Sep 14, 2025

Existing Transformer-based text-to-speech (TTS) systems commonly employ Rotary Position Embedding (RoPE), which relies on absolute token indices and thus struggles to model the dynamic alignment between text and speech—particularly under variable speech durations or long-sequence generation, leading to significant performance degradation. To address this, we propose Length-Aware Rotary Position Encoding (LARoPE), which explicitly encodes relative query-key distances via length-normalized positional indices, thereby enhancing robustness in cross-modal alignment. Integrated into the Transformer’s cross-attention mechanism, LARoPE requires no additional parameters while improving stability in long-sequence modeling. Experiments demonstrate that LARoPE accelerates convergence, improves alignment accuracy and speech naturalness, achieves state-of-the-art word error rate on zero-shot TTS benchmarks, and enables high-fidelity synthesis of utterances up to 30 seconds in duration.

0 citationsRead paper