Raon-OpenTTS-Core: Curated TTS Training Data

• Raon-OpenTTS-Core is a model-filtered subset of Raon-OpenTTS-Pool designed to boost TTS quality by excluding noisy, misaligned, and off-domain segments.
• It uses a five-stage filtering pipeline—including VAD, ASR consistency, speaker homogeneity, and noise robustness—to ensure high transcript and acoustic fidelity.
• The curated Core dataset underpins DiT-based TTS model training, yielding improved WER and speaker similarity compared to models trained on the full Pool.

Raon-OpenTTS-Core is a curated, high-quality subset of Raon-OpenTTS-Pool, introduced in the paper "Raon-OpenTTS: Open Models and Data for Robust Text-to-Speech" (Kim et al., 20 May 2026). It was created to support large-scale, fully open, reproducible English text-to-speech training by retaining broad source coverage while removing segments likely to degrade end-to-end TTS quality, including noisy, mis-aligned, or off-domain utterances. Within the Raon-OpenTTS framework, Core is the dataset on which the released diffusion transformer (DiT)-based TTS models are trained, and it is positioned as the dataset component that converts a very large mixed-source collection into a more uniform training substrate (Kim et al., 20 May 2026).

1. Definition and motivation

Raon-OpenTTS-Core is defined as a model-filtered subset of the larger Raon-OpenTTS-Pool. Raon-OpenTTS-Pool was assembled as a fully open, reproducible collection of English speech data for large-scale TTS training, with 615,000 hours and 240,000,000 segments. The Core subset reduces this to 510,000 hours and 194,000,000 segments through a multi-stage filtering pipeline (Kim et al., 20 May 2026).

The stated motivation is that raw public and web-sourced collections inevitably contain noisy, mis-aligned, or off-domain utterances, including poor transcripts, non-speech sounds, and cross-speaker contamination. The purpose of Core is therefore not merely dataset reduction, but selective retention of segments that preserve synthesis fidelity by limiting artifacts in synthesis, speaker drift, and unnatural prosody (Kim et al., 20 May 2026).

The reduction from Pool to Core is described as a approximately 17% reduction in overall size, with a larger improvement in per-segment fidelity. This framing is technically significant because it separates raw scale from usable scale. A plausible implication is that, in this pipeline, corpus quality control is treated as a first-order determinant of TTS robustness rather than as a secondary cleanup step.

2. Composition of the source pool

Raon-OpenTTS-Core inherits its provenance from Raon-OpenTTS-Pool, which aggregates standard public TTS and ASR corpora together with web-sourced recordings. The public corpora listed for Pool are LibriSpeech, CMU Arctic, L2-Arctic, CSTR VCTK, TED-LIUM 3, AMI Meeting, Switchboard, Common Voice, GigaSpeech, VoxPopuli, and The People’s Speech (Kim et al., 20 May 2026).

The web-sourced recordings consist of YODAS, described as YouTube-oriented scraped segments, and various CC-BY podcasts, lecture clips, and open interviews (Kim et al., 20 May 2026). In aggregate, Pool contains 615,000 hours of speech and 240,000,000 segments.

This mixed-source composition matters because it explains both the scale and the need for stringent filtering. Public corpora and web-sourced recordings expand coverage substantially, but they also introduce heterogeneity in transcript quality, speaker purity, background noise, and acoustic domain. This suggests that Core is best understood not as a separately collected dataset, but as the output of a normalization and curation process applied to a deliberately diverse upstream pool.

3. Filtering pipeline

Raon-OpenTTS-Core is derived through a five-stage filtering recipe in which segments below specified thresholds are discarded (Kim et al., 20 May 2026). The stages are explicit and reproducible.

First, duration and metadata sanity filtering removes segments shorter than 0.5 s or longer than 15 s, and discards segments with non-ASCII or non-printable characters in the transcript. This stage constrains gross temporal outliers and transcript encoding anomalies (Kim et al., 20 May 2026).

Second, a neural VAD model is applied to each segment to compute the speech-to-total time ratio

$$f_{\mathrm{VAD}}=\frac{\text{total speech frames}}{\text{total frames}}.$$

Only segments with $f_{\mathrm{VAD}} \geq 0.65$ are kept. The paper states that this ensures low music and crosstalk (Kim et al., 20 May 2026).

Third, ASR consistency filtering re-transcribes each segment using OWSM v3.1 and computes word error rate against the provided transcript:

$\mathrm{WER}=\frac{S+D+I}{N}.$

The derived score is defined as $f_{\mathrm{ASR}} = 1 - \mathrm{WER}$, and the threshold is $f_{\mathrm{ASR}} \geq 0.80$, equivalently $\mathrm{WER} \leq 20\%$ (Kim et al., 20 May 2026). This stage operationalizes transcript reliability.

Fourth, speaker homogeneity filtering extracts per-segment speaker embeddings via a Res2Net-based encoder. For each hypothesized speaker cluster, the centroid $c$ is computed, and each segment embedding $e$ is scored by cosine similarity,

$f_{\mathrm{spk}}=\mathrm{cosine}(e,c).$

Only segments with $f_{\mathrm{spk}} \geq 0.75$ are retained, filtering multi-speaker or off-voice segments (Kim et al., 20 May 2026).

Fifth, an optional noise robustness stage runs DNSMOS on each segment to estimate perceptual noise score $f_{\mathrm{VAD}} \geq 0.65$0, keeping only segments with $f_{\mathrm{VAD}} \geq 0.65$1 (Kim et al., 20 May 2026).

The paper also reports an early composite scoring formulation,

$f_{\mathrm{VAD}} \geq 0.65$2

with $f_{\mathrm{VAD}} \geq 0.65$3 and the requirement $f_{\mathrm{VAD}} \geq 0.65$4. However, the final release adopts cascading thresholds instead, described as more transparent, nearly as selective, and easier to reproduce (Kim et al., 20 May 2026).

A concise summary of the final filtering stages is given below.

Stage	Signal	Threshold
1	Duration and metadata sanity	Keep 0.5 s to 15 s; discard non-ASCII or non-printable transcript characters
2	Voice activity detection	$f_{\mathrm{VAD}} \geq 0.65$5
3	ASR consistency	$f_{\mathrm{VAD}} \geq 0.65$6; equivalently WER $f_{\mathrm{VAD}} \geq 0.65$7
4	Speaker homogeneity	$f_{\mathrm{VAD}} \geq 0.65$8
5	Noise robustness (optional stage)	$f_{\mathrm{VAD}} \geq 0.65$9

The methodological importance of the pipeline lies in its decomposition of quality into transcript alignment, speech occupancy, speaker purity, and acoustic cleanliness. Rather than relying on a single scalar quality score in the final release, the pipeline preserves interpretability by making each exclusion criterion explicit.

4. Final dataset statistics and quality profile

After filtering, Raon-OpenTTS-Core contains 510,000 hours of speech and 194,000,000 segments (Kim et al., 20 May 2026). Relative to Pool, the paper reports the following quality metrics for Core: median WER (ASR check) of 5.3% versus 12.1% for Pool, mean VAD ratio of 0.92 versus 0.79, mean $\mathrm{WER}=\frac{S+D+I}{N}.$0 of 4.21 versus 3.45, and mean signal-to-noise ratio via STFT estimate of 23.4 dB versus 17.8 dB (Kim et al., 20 May 2026).

These differences indicate that filtering materially changes the distributional profile of the training data. The lower median WER suggests tighter transcript-speech consistency; the higher mean VAD ratio indicates denser speech occupancy; the higher DNSMOS and SNR values indicate cleaner acoustics. A plausible implication is that Core is optimized for reducing supervisory noise across both linguistic and acoustic dimensions, rather than for maximizing raw speaker or domain diversity.

An important misconception is that the larger upstream pool should necessarily be superior because it contains more hours. The reported statistics do not support that interpretation. In the Raon-OpenTTS setting, the filtered subset has fewer hours yet stronger quality indicators on every metric explicitly reported for comparison (Kim et al., 20 May 2026).

5. Role in training the Raon-OpenTTS models

Raon-OpenTTS-Core served as the sole training data for a family of DiT-based TTS models. The paper reports two trained model sizes, a 0.3B configuration and a 1.0B configuration, with the following architecture parameters (Kim et al., 20 May 2026).

Config	Layers / Heads / EmbDim / FF Dim	Params
0.3B	22 / 16 / 1024 / 2048	336 M
1.0B	28 / 22 / 1408 / 5632	1048 M

The training recipe is identical for both sizes except for batch scaling: AdamW with $\mathrm{WER}=\frac{S+D+I}{N}.$1, $\mathrm{WER}=\frac{S+D+I}{N}.$2, and weight decay $\mathrm{WER}=\frac{S+D+I}{N}.$3; warm-up for the first 10 K steps; cosine annealing thereafter over 200 K steps total; and a batch of 32 segments per GPU across 256 A100 GPUs (Kim et al., 20 May 2026). The loss is $\mathrm{WER}=\frac{S+D+I}{N}.$4 on mel-spectrogram plus a KL term for the variance predictor. The vocoder is HiFi-GAN v3, trained separately on Core until MOS $\mathrm{WER}=\frac{S+D+I}{N}.$5 (Kim et al., 20 May 2026).

This training setup places Core at the center of the system’s reproducibility claim. Because the same open dataset underlies both the acoustic model and the separately trained vocoder, Core functions as the canonical data substrate for the released model family. This suggests that Core is not merely a benchmark-oriented subset, but the intended production-grade training corpus within the open pipeline.

6. Empirical impact on TTS performance

The paper reports direct comparisons between models trained on Core and otherwise identical models trained on Pool. On Seed-TTS-Eval, the Raon-OpenTTS-1B model trained on Core achieves WER = 1.78% and SIM = 0.749, whereas the same model trained on Pool achieves WER = 2.03% and SIM = 0.730 (Kim et al., 20 May 2026). On CV3-Hard-EN, the Core-trained model achieves WER = 6.15% and SIM = 0.775, compared with WER = 6.85% and SIM = 0.743 for the Pool-trained version (Kim et al., 20 May 2026).

The paper also reports filtering-stage ablations. Dropping the speaker filter while keeping the others raises WER on Seed-TTS-Eval to 1.98% and reduces SIM to 0.715. Dropping the ASR filter inflates WER to 2.34% and reduces SIM to 0.698 (Kim et al., 20 May 2026). These results are explicitly interpreted as confirming that each filter contributes meaningfully to final fidelity.

In the broader evaluation summarized in the paper, Raon-OpenTTS-1B is described as showing comparable performance to state-of-the-art models such as Qwen3-TTS and CosyVoice 3, despite those systems being trained on several million hours of proprietary speech data (Kim et al., 20 May 2026). On Seed-TTS-Eval, it ranks second on WER and first on SIM among recent open-weight TTS baselines; on CV3-Hard-EN, it ranks first on both metrics; and on Raon-OpenTTS-Eval, it achieves the best average WER and SIM among all evaluated models and the second-best human preference as measured by CMOS (Kim et al., 20 May 2026).

The significance of Core in these results is direct rather than incidental: the comparison against Pool isolates the effect of data curation. The evidence presented does not merely show that open-data TTS can work at scale; it shows that model-based filtering of open data can materially improve robustness and speaker similarity even when the starting pool is already very large.

7. Operational considerations, limitations, and prospective extensions

The paper identifies computational cost as the main bottleneck of the filtering process, stating that large-scale ASR and speaker-embedding inference require approximately 150 GPU-days for 615 K hours (Kim et al., 20 May 2026). The recommended operational practice is to stream the pipeline and checkpoint intermediate results.

A further practical condition is transcript availability. The filtering procedure requires access to ground-truth transcripts, which the paper describes as common for public corpora and harder for web data (Kim et al., 20 May 2026). For truly unsupervised collections, the proposed substitute is round-trip ASR$\mathrm{WER}=\frac{S+D+I}{N}.$6TTS consistency.

The paper also identifies future improvements: dynamic thresholds via a small held-out validation set rather than fixed WER and VAD cuts; extension of the pipeline to multi-language or code-switched data; and incorporation of prosody or emotion classifiers to preserve expressiveness (Kim et al., 20 May 2026). These are prospective rather than realized components.

Taken together, these points delimit both the strengths and the boundaries of Raon-OpenTTS-Core. It is a large open English TTS training corpus shaped by explicit, model-based quality controls, but it remains dependent on transcript-bearing data sources and computationally intensive filtering. A plausible implication is that its methodology is portable, while its exact operating point reflects a trade-off among reproducibility, filtering cost, and fidelity targets specific to the Raon-OpenTTS training regime.