Supertonic: Robust Lightweight TTS Model

• Supertonic is an open-weight, lightweight text-to-speech model designed to deliver robust, high-quality synthesis from noisy, in-the-wild data.
• It integrates a compact speech autoencoder, a flow-matching text-to-latent generator, and an utterance-level duration predictor for precise text-speech alignment.
• The implementation of Self-Purifying Flow Matching (SPFM) effectively mitigates label noise, enhancing intelligibility and reducing word error rates.

Supertonic is an open-weight, lightweight text-to-speech (TTS) model architecture designed to facilitate robust and scalable speech synthesis, particularly when trained on unconstrained, noisy in-the-wild data. Its fundamental design comprises a compact speech autoencoder, a flow-matching-based text-to-latent generator, and an utterance-level duration predictor, all tightly coupled within a unified system. Supertonic, when combined with Self-Purifying Flow Matching (SPFM), attains state-of-the-art robustness and intelligibility under adverse training conditions, as demonstrated in the WildSpoof 2026 TTS Track (Yi et al., 19 Dec 2025).

1. System Architecture

Supertonic's architecture consists of three interacting modules:

1. Speech Autoencoder: The encoder $E$ transforms an input waveform $x$ or its mel-spectrogram into a sequence of continuous latent variables $z = E(x)$. The decoder $D$ reconstructs the waveform $\hat{x} = D(z)$, with the autoencoder trained a priori to minimize $\|\hat{x} - x\|$, ensuring a compact and smooth latent space.
2. Duration Predictor: Given a text input $c$ (e.g., phoneme sequence), the duration predictor generates per-token durations $d_1,\ldots,d_n$, which are used to expand $c$ to a frame-level sequence $c'$ aligned with the temporal granularity of the latent representation.
3. Flow-Matching Text-to-Latent Generator: The model conditions on $x$0 to generate a target latent sequence $x$1 from Gaussian noise $x$2 via linear interpolation:

$x$3

A neural network $x$4 predicts the “velocity” $x$5. For synthesis, the ODE

$x$6

is solved to yield $x$7, which is decoded to waveform space via $x$8.

2. Training Objectives and Self-Purifying Flow Matching (SPFM)

The classical flow-matching framework optimizes the conditional loss:

$x$9

and, for classifier-free guidance (CFG), the unconditional loss:

$z = E(x)$0

SPFM introduces an explicit mechanism for label noise mitigation. After a 1,000-step warmup, per-sample conditional and unconditional losses

$z = E(x)$1

are computed. The guiding hypothesis is that for well-aligned $z = E(x)$2 pairs,

$z = E(x)$3

Samples are then routed:

$z = E(x)$4

Flagged (potentially noisy) samples update only the unconditional branch, retaining acoustic diversity but avoiding spurious conditional associations.

3. Implementation Workflow and Experimental Protocol

• Datasets: Fine-tuning is conducted on TITW-easy and TITW-hard subsets of the WildSpoof TTS in-the-wild dataset, with mini-batches sampling a 1:1 ratio from each subset.
• Optimization: The model is fine-tuned for 10,000 steps with a batch size of 32 on four NVIDIA RTX A100 GPUs.
• SPFM Application: SPFM is activated after the initial 1,000-step warmup to minimize false positive routing. The interpolation parameter $z = E(x)$5 is primarily sampled near 0.5 for stability.
• Validation Splits: Evaluation covers four canonical splits—Known-Speaker Known-Text (KSKT), Known-Speaker Unknown-Text (KSUT), Unknown-Speaker Known-Text (USKT), and Unknown-Speaker Unknown-Text (USUT)—supplemented by Librispeech/VoxCeleb-derived subsets.

4. Evaluation Metrics and Performance

Supertonic with SPFM is assessed using multiple quantitative and perceptual metrics:

Metric	KSKT	KSUT–USUT Range	WS Seen	WS Unseen
WER (%)	3.26	4.75–6.53	5.50	5.88
CER (%)	2.33	—	—	—
UTMOS	3.58	3.57–4.03	3.7390	3.9078
DNSMOS	2.96	2.96–3.19	3.0780	3.1195
Spk-sim	0.59	—	—	—
MCD (dB)	8.59	—	—	—

The system achieved the lowest Word Error Rate (WER) among all teams in the WildSpoof 2026 TTS Track, while ranking second in perceptual metrics such as UTMOS and DNSMOS. On unseen speakers, it achieved the highest UTMOS and second-best DNSMOS (Yi et al., 19 Dec 2025).

5. Qualitative Phenomena and Robustness

Analysis reveals strong robustness against noise and reverberation, with utterances containing background chatter or intense room effects yielding minimal transcription errors. Alignment between text and speech is highly consistent, with pronounced reductions in mispronunciations and word omissions relative to baselines fine-tuned naively on noisy data. Generated speech embeddings (x-vectors) display close speaker similarity to reference signals, including for out-of-domain voices.

6. Methodological Significance and Extensibility

Supertonic, augmented by SPFM, exemplifies that lightweight, open-weight TTS models can be robustly adapted to diverse, unconstrained speech data without the need for aggressive data pre-filtering. SPFM constitutes a general recipe: it is applicable to any conditional flow-matching or diffusion-based model for robust self-cleaning in the presence of noisy labels (Yi et al., 19 Dec 2025). Prospective research directions include adaptive thresholding of the conditional/unconditional loss gap, curriculum-driven warmup periods, extension to video-to-speech or other modalities, multilingual scaling, style and emotion conditioning, and integration with front-end ASR feedback loops.

7. Broader Implications

The Supertonic system, particularly when trained with Self-Purifying Flow Matching, demonstrates that explicit, in-training noise detection mechanisms enable scalable speech synthesis with high intelligibility and perceptual quality, even under “wildest, noisiest” conditions. This suggests a generalizable paradigm for robust TTS applicable to real-world, open-domain environments (Yi et al., 19 Dec 2025).