Raon-OpenTTS-Pool: Open TTS Speech Dataset

• Raon-OpenTTS-Pool is a massive English speech dataset aggregated from diverse public sources, featuring approximately 615,000 hours of audio and 240 million speech–text segments.
• A multi-stage model-based filtering pipeline refines the dataset into Raon-OpenTTS-Core, significantly improving intelligibility (from 0.92 to 0.98) and quality (MOS from 3.2 to 4.1).
• It enables reproducible TTS research by providing open data, training recipes, and checkpoints for diffusion transformer models ranging from 0.3B to 1B parameters.

Searching arXiv for the specified paper to ground the article in the cited source. Raon-OpenTTS-Pool is a large-scale open English speech dataset introduced alongside the Raon-OpenTTS text-to-speech system in "Raon-OpenTTS: Open Models and Data for Robust Text-to-Speech" (Kim et al., 20 May 2026). It is designed for reproducible TTS training at a scale that had often been associated with proprietary corpora. The dataset consists of approximately 615,000 hours of audio and approximately 240 million speech–text segments, aggregated from publicly available English speech corpora and web-sourced recordings. A model-based filtering pipeline is applied to this pool to derive Raon-OpenTTS-Core, a curated subset of approximately 510,000 hours and approximately 194 million segments, which is then used to train diffusion transformer (DiT)-based TTS models from 0.3B to 1B parameters (Kim et al., 20 May 2026).

1. Definition, scope, and composition

Raon-OpenTTS-Pool is an English-only speech dataset assembled from heterogeneous open sources and web data for large-scale TTS research (Kim et al., 20 May 2026). Its stated total size is approximately 615,000 hours of audio and approximately 240 million speech–text segments. The major contributors include Librispeech (960 h), Common Voice (approximately 9,000 h), GigaSpeech (approximately 10,000 h), VoxPopuli (approximately 4,000 h), TED-LIUM 3, VCTK, CMU Arctic, L2-ARCTIC, AMI, Switchboard, The People’s Speech, and approximately 200,000 hours of YouTube-sourced recordings.

The dataset spans a broad range of acoustic conditions. These include clean studio reads, telephone/meeting speech, in-the-wild YouTube videos, expressive/emotional material such as CREMA-D and Emotional Voices Database, noisy environments, and multi-channel recordings (Kim et al., 20 May 2026). This heterogeneity is central to the stated purpose of supporting robust TTS training rather than narrowly optimized synthesis under laboratory conditions.

Each segment is accompanied by structured metadata. The per-segment fields are transcript, speaker ID, source corpus, sampling rate, estimated SNR, and a condition label drawn from the set $$\{\text{clean}, \text{noisy}, \text{in-the-wild}, \text{expressive}\}$$ (Kim et al., 20 May 2026). This metadata design links data provenance, acoustic characterization, and speaker identity, which is directly relevant to reproducible corpus construction and controlled evaluation.

A plausible implication is that Raon-OpenTTS-Pool is not merely a raw aggregation of corpora, but a dataset organized to support both model training and post hoc analysis across speaker, source, and condition axes.

2. Data curation and filtering pipeline

Raon-OpenTTS-Pool is transformed into Raon-OpenTTS-Core through a multi-stage filtering pipeline intended to remove unintelligible, low-quality, or speaker-ambiguous segments (Kim et al., 20 May 2026). The pipeline is explicitly model-based and operates on each audio segment $x_i$ with transcript $y_i$.

The first stage is intelligibility scoring. Word error rate is computed via a robust ASR system such as Whisper:

$\mathrm{WER}(x_i) = \frac{S_i + D_i + I_i}{N_i}$

where $S_i$, $D_i$, and $I_i$ denote substitution, deletion, and insertion counts, and $N_i$ is the reference word count. An intelligibility score is then defined as

$S_{\mathrm{int}}(x_i) = 1 - \mathrm{WER}(x_i).$

The threshold is

$$S_{\mathrm{int}}(x_i)\ge 0.70 \quad\Longleftrightarrow\quad \mathrm{WER}(x_i)\le 0.30.$$

The second stage is speech-quality scoring. A non-intrusive MOS predictor such as DNSMOS is used to produce

$x_i$0

and segments are retained only if

$x_i$1

The third stage addresses speaker consistency through clustering. Fixed-dimensional speaker embeddings $x_i$2 are extracted, for example with ECAPA-TDNN, and pairwise distance is defined as

$x_i$3

Agglomerative clustering is then performed over $x_i$4. Clusters are discarded if they contain fewer than 20 segments or if their average intra-cluster distance exceeds 0.6:

$x_i$5

The fourth stage enforces duration and text-audio alignment consistency. The ratio

$x_i$6

is computed, and only segments satisfying

$x_i$7

seconds per character are kept. The data summary states that this empirically covers natural speaking rates (Kim et al., 20 May 2026).

This pipeline operationalizes four distinct desiderata: transcript intelligibility, perceived signal quality, speaker identity coherence, and plausible alignment between textual length and acoustic duration. The explicit thresholds make the curation process inspectable and reproducible.

3. Raon-OpenTTS-Core and post-filtering characteristics

After the four filtering stages, approximately 510,000 hours and approximately 194 million segments remain, forming Raon-OpenTTS-Core (Kim et al., 20 May 2026). This corresponds to 83% of the original pool by hour. Raon-OpenTTS-Core is the subset used for model training.

The paper reports quality improvements on a held-out subset in mean $x_i$8 standard deviation form. Intelligibility improves from $x_i$9 $y_i$0, the MOS predictor score improves from $y_i$1 $y_i$2, and ASR WER decreases from $y_i$3 (Kim et al., 20 May 2026). These values provide an empirical summary of the effect of filtering, rather than merely documenting threshold choices.

The transition from Pool to Core is therefore not only a reduction in corpus size but also a measurable redistribution toward higher intelligibility and higher predicted quality. This suggests that the filtering pipeline functions as a quality-concentrating transformation while preserving most of the total training hours.

Because the article’s source describes Raon-OpenTTS-Core as a curated, high-quality subset, it is appropriate to interpret it as the canonical training partition within the Raon-OpenTTS ecosystem rather than a secondary convenience split.

4. Statistical structure of the corpus

The dataset summary reports several corpus-level statistics for segment durations and speaker structure (Kim et al., 20 May 2026). Let $y_i$4 denote the segment durations. The empirical duration distribution $y_i$5 has mean $y_i$6 s, median $y_i$7 s, and range $y_i$8 s. The duration histogram is summarized as

$y_i$9

These figures indicate that the majority of segments fall in the 3-to-10-second regime, while both shorter and longer utterances remain substantially represented. For TTS training, that balance is relevant because it avoids collapsing the corpus into either very short prompt-like segments or predominantly long-form recordings.

Speaker structure is also explicitly characterized. The corpus contains approximately $\mathrm{WER}(x_i) = \frac{S_i + D_i + I_i}{N_i}$0 inferred speaker clusters, and the distribution has a power-law tail in which the top 1% of speakers account for 12% of total hours (Kim et al., 20 May 2026). Per-speaker segment count has mean approximately 850 segments, and 10% of speakers have more than 5,000 segments.

These statistics point to a large and highly imbalanced speaker inventory, which is typical of web-scale and aggregated corpora. A plausible implication is that the dataset can support both broad multi-speaker modeling and analyses of long-tail speaker coverage, although any claim about downstream fairness or representation would require evidence beyond the reported summary statistics.

5. Role in reproducible TTS training

Raon-OpenTTS-Pool is framed as infrastructure for reproducible large-scale TTS training (Kim et al., 20 May 2026). The published resources include the full Raon-OpenTTS-Pool, the filtering code, and training recipes for DiT-based models ranging from 0.3B to 1B parameters. The released checkpoints include open Raon-OpenTTS models at 0.3B and 1B parameters trained on Raon-OpenTTS-Core.

Within this training setup, Raon-OpenTTS-Core functions as the filtered data substrate for the Raon-OpenTTS model family. The associated paper states that Raon-OpenTTS is a series of diffusion transformer-based TTS models from 0.3B to 1B parameters, trained using the curated core subset (Kim et al., 20 May 2026). This creates a fully open chain linking raw data, curation, training recipe, and released checkpoints.

The importance of this arrangement lies in reproducibility. By publishing the raw pool, the curated core, the filtering pipeline, the model code, and the checkpoints, the project enables end-to-end replication and controlled variation of corpus preparation and model training (Kim et al., 20 May 2026). In TTS research, this matters because performance differences are often confounded by inaccessible training data. The open release reduces that confound for this particular training stack.

This suggests that Raon-OpenTTS-Pool is intended not only as a static benchmark resource but also as a baseline-generating substrate for subsequent open TTS experiments.

6. Evaluation framework and benchmark coupling

The dataset is paired with a structured evaluation suite, Raon-OpenTTS-Eval, to assess TTS robustness across four acoustic conditions: clean, noisy, in-the-wild, and expressive (Kim et al., 20 May 2026). The benchmark uses ASR-based WER and speaker similarity (SIM) as objective metrics, together with subjective evaluation via SMOS and CMOS.

WER is defined in the benchmark exactly as

$\mathrm{WER}(x_i) = \frac{S_i + D_i + I_i}{N_i}$1

Speaker similarity is defined as

$\mathrm{WER}(x_i) = \frac{S_i + D_i + I_i}{N_i}$2

Subjective evaluation includes SMOS on a 5-point scale for speaker/style similarity and CMOS on a $\mathrm{WER}(x_i) = \frac{S_i + D_i + I_i}{N_i}$3 scale for pairwise quality (Kim et al., 20 May 2026). The evaluation splits are specified as 30 items with 6 raters per condition.

The benchmark coupling is significant because the same condition labels that appear in dataset metadata—clean, noisy, in-the-wild, and expressive—also structure evaluation. This creates alignment between training data annotation and robustness assessment. A plausible implication is that the benchmark is designed to expose whether models trained on heterogeneous open data retain performance under acoustically diverse conditions, rather than reporting only aggregate synthesis quality on homogeneous test sets.

The source further reports model-level outcomes: on Raon-OpenTTS-Eval, Raon-OpenTTS-1B 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). In this context, Raon-OpenTTS-Pool is not separable from the evaluation apparatus; together they define a reproducible training-and-measurement environment.

7. Reported performance and research significance

The Raon-OpenTTS paper positions the dataset as enabling open models to approach the performance of closed-data systems (Kim et al., 20 May 2026). Using Raon-OpenTTS-Core, the 1B-parameter model is reported to perform comparably to state-of-the-art models such as Qwen3-TTS and CosyVoice 3, which are described as being trained on several million hours of proprietary speech data.

Specific benchmark results are provided. On Seed-TTS-Eval, Raon-OpenTTS-1B achieves a WER of 1.78% and a speaker similarity of 0.749, ranking second on WER and first on SIM among recent open-weight TTS baselines (Kim et al., 20 May 2026). On CV3-Hard-EN, it achieves a WER of 6.15% and a SIM of 0.775, ranking first on both metrics. On Raon-OpenTTS-Eval, it achieves the best average WER and SIM among all evaluated models, together with the second-best human preference under CMOS.

These claims concern the trained model rather than the dataset in isolation, but they are central to the significance of Raon-OpenTTS-Pool because the dataset is presented as the enabling open-data substrate. The principal research claim is therefore not merely that a large open corpus exists, but that a corpus of this design and scale can support competitive TTS training under fully open and reproducible conditions (Kim et al., 20 May 2026).

A common misconception in large-scale generative modeling is that performance parity with closed systems necessarily requires closed training data. The evidence reported here does not eliminate all confounds, but it directly contests that assumption within the scope of English TTS by combining an open 615,000-hour pool, a transparent filtering pipeline, open DiT training recipes, and public checkpoints (Kim et al., 20 May 2026).