arXiv · 2025 · Preprint
Yi, Kim & Lee (Supertone Inc.) · → Paper · Demo: ? · Code: ?
A WildSpoof 2026 TTS Track challenge submission that fine-tunes the lightweight, flow-matching-based Supertonic TTS model with Self-Purifying Flow Matching (SPFM), a per-sample noise-detection training mechanism, achieving the best word error rate among all participating teams on noisy, in-the-wild training and evaluation data.
Problem
TTS research traditionally relies on clean, studio-recorded, carefully curated training data, which limits scalability but enables stable text-speech alignment. In-the-wild speech, with background noise, reverberation, device variability, and inconsistent or outright incorrect labeling, offers a more realistic but much harder training regime. The WildSpoof Challenge 2026 TTS Track specifically benchmarks TTS systems trained under these unconstrained conditions, evaluating intelligibility (ASR-measured WER), perceptual quality (UTMOS, DNSMOS), and faithfulness (speaker similarity, Mel Cepstral Distance). Building a system that performs reliably here requires explicitly handling label noise, unpredictable duration variation, and degraded text-speech alignment, challenges that conventional flow-matching TTS pipelines are not designed to handle.
Method
The submission fine-tunes Supertonic, a lightweight, pre-existing flow-matching TTS architecture (a speech autoencoder for continuous latent representation, a flow-matching text-to-latent generator, and an utterance-level duration predictor), using Self-Purifying Flow Matching (SPFM), a noise-mitigation training mechanism from the same authors’ prior work. SPFM operates within the classifier-free-guidance framing of conditional flow matching: for each text-speech pair, it computes both a conditional flow-matching loss (conditioned on the text label) and an unconditional flow-matching loss (with conditioning dropped) at the same sampled interpolation point along the flow trajectory. The key intuition is that if a text label is correct, the conditional loss should not exceed the unconditional loss in expectation; when a sample’s conditional loss does exceed its unconditional loss, SPFM treats the label as potentially unreliable and routes that sample to unconditional-only training for that step, still extracting useful acoustic information from it without letting a wrong label corrupt the conditional objective. SPFM is activated only after a 1,000-step warm-up (to avoid spurious detections while the model is still undertrained) and uses a fixed interpolation time near the trajectory midpoint. The system was fine-tuned for 10,000 iterations (batch size 32, four A100 GPUs) on the challenge-provided TITW-easy and TITW-hard subsets, sampled in a 1:1 ratio to balance relatively clean and noisy conditions.
Key Results
On four internal validation sets spanning known/unknown speakers and known/unknown text, the system achieves a 3.26% WER on the cleanest condition (KSKT) with stable perceptual quality across all conditions (UTMOS 3.578-4.029, DNSMOS 2.962-3.193) and reasonably preserved speaker similarity even for unknown speakers. On the official WildSpoof 2026 TTS Track leaderboard (Team T02, evaluated against six other teams), the system achieved the lowest WER of any team for both seen (5.50%) and unseen (5.88%) speaker conditions, well ahead of several competing systems whose WER exceeded 20% or, in one case, 99%. The system ranked second in perceptual quality (UTMOS, DNSMOS) behind the top team, though the gap was small, and its unseen-speaker UTMOS score (3.9078) was the highest of any submission.
Novelty Assessment
Neither SPFM (introduced in a separate prior paper by overlapping authors) nor Supertonic (also from prior work) is new to this paper; the contribution is applying and validating this specific combination in a genuinely adversarial, multi-team competitive setting with real noisy in-the-wild data, rather than proposing a new architecture or training technique. As a two-page challenge-track system description, its value lies in the concrete empirical demonstration that a compact, efficient TTS architecture combined with an explicit per-sample noise-detection mechanism can outperform larger or more complex competing systems on intelligibility under realistic label-noise conditions.
Field Significance
moderate — This is a focused, real-world validation of an existing noise-robust training technique in a genuinely competitive, adversarial benchmark setting (best WER among seven teams), rather than a new methodological contribution. Its main value to the field is demonstrating that lightweight, efficient TTS architectures do not need to sacrifice robustness to label noise when paired with an appropriate training-time noise-detection mechanism, a practically relevant data point for anyone deploying compact TTS systems on uncurated, in-the-wild data.
Claims
- supports: A per-sample comparison between a conditional and unconditional flow-matching loss, computed at the same interpolation point during training, can identify unreliable text-speech pairs in noisy in-the-wild TTS training data without a separate label-verification model or manual filtering.
Evidence: SPFM routes a sample to unconditional-only training whenever its conditional flow-matching loss exceeds its unconditional loss at the same interpolation time, still extracting acoustic information from suspicious samples rather than discarding them entirely. (§2.2)
- supports: A compact, lightweight flow-matching TTS architecture, fine-tuned with an explicit per-sample noise-detection training mechanism, can achieve better intelligibility on noisy in-the-wild data than substantially larger or more complex competing systems.
Evidence: The SPFM-augmented Supertonic system achieved the lowest WER among all seven participating teams in the WildSpoof 2026 TTS Track official leaderboard, for both seen (5.50%) and unseen (5.88%) speaker conditions, while several competing teams’ WER exceeded 20% or even 90%. (§2.4, Tables 2-3)
- complicates: Optimizing a TTS system for robustness to noisy training labels does not guarantee the single best result on every evaluation dimension simultaneously; a system can lead on intelligibility while trailing slightly on perceptual quality metrics.
Evidence: The SPFM-augmented Supertonic system achieved the best WER among all teams but ranked second on UTMOS/DNSMOS perceptual quality metrics, though the gap to the top-ranked team was small and its unseen-speaker UTMOS score was the highest among all submissions. (§2.4, Tables 2-3)
Limitations and Open Questions
As a two-page challenge-submission paper, the work does not include ablations isolating SPFM’s specific contribution relative to fine-tuning Supertonic without it, nor analysis of what fraction of training samples SPFM routes to unconditional training or how that fraction evolves over training. Results are reported only on the WildSpoof 2026 challenge’s specific datasets and evaluation protocol, leaving open how the approach generalizes to other in-the-wild TTS domains, languages, or noise conditions beyond the challenge’s own data.
Wiki Connections
- Flow Matching — fine-tunes a flow-matching-based TTS architecture (Supertonic) using a per-sample noise-detection mechanism that compares conditional and unconditional flow-matching losses within the classifier-free-guidance framing.
- Zero-Shot TTS — evaluated on genuinely unseen speakers (US conditions) with real speaker-similarity results alongside seen-speaker performance.