Raon-OpenTTS: Open Models and Data for Robust Text-to-Speech

Abstract: Recent advances in text-to-speech (TTS) models show impressive speech naturalness and quality, yet the role of large-scale open data in driving this progress remains underexplored. In this work, we introduce Raon-OpenTTS, an open TTS model that performs competitively with state-of-the-art closed-data TTS models, and Raon-OpenTTS-Pool, a large-scale open dataset for reproducible TTS training. Raon-OpenTTS-Pool consists of 615K hours of 240M speech segments aggregated from publicly available English speech corpora and web-sourced recordings. With a model-based filtering pipeline applied to Raon-OpenTTS-Pool, we derive Raon-OpenTTS-Core, a curated, high-quality subset of 510K hours and 194M speech segments. Using Raon-OpenTTS-Core, we train Raon-OpenTTS, a series of diffusion transformer (DiT)-based TTS models from 0.3B to 1B parameters. On multiple benchmarks, Raon-OpenTTS-1B shows comparable performance to state-of-the-art models such as Qwen3-TTS and CosyVoice 3, which are trained on several million hours of proprietary speech data. Notably, on Seed-TTS-Eval, Raon-OpenTTS-1B achieves a word error rate (WER) of 1.78% and a speaker similarity (SIM) of 0.749, ranking second on WER and first on SIM among recent open-weight TTS baselines. On CV3-Hard-EN, Raon-OpenTTS-1B achieves a WER of 6.15% and a SIM of 0.775, ranking first on both metrics. Furthermore, to support robust evaluation, we introduce Raon-OpenTTS-Eval, a structured benchmark for assessing TTS robustness across diverse acoustic conditions including clean, noisy, in-the-wild, and expressive speech. 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 comparative mean opinion score (CMOS). Our data pool, filtering pipeline, training code, and checkpoints are publicly available at https://github.com/krafton-ai/RAON-OpenTTS.

• The paper introduces Raon-OpenTTS—a zero-shot TTS model series that uses large-scale, curated open-access speech data and advanced filtering to achieve state-of-the-art performance.
• It employs a DiT-based architecture with log-mel spectrogram encoding and efficient training on NVIDIA hardware, yielding competitive metrics on WER and speaker similarity.
• The study demonstrates that rigorous data curation and scaling can enable open-source TTS models to match or exceed proprietary systems, promoting reproducible research.

Raon-OpenTTS: An Open-Data, Open-Weight Framework for Robust Text-to-Speech

Overview

This work introduces Raon-OpenTTS, a family of zero-shot text-to-speech (TTS) models leveraging a newly curated, large-scale, open-access speech corpus—Raon-OpenTTS-Pool—and a robust data filtering pipeline culminating in Raon-OpenTTS-Core. The models, employing a DiT-based architecture, achieve competitive performance relative to the latest proprietary-data-trained SOTA systems. The contribution includes not only models and data but also Raon-OpenTTS-Eval, a comprehensive evaluation benchmark spanning varied acoustic and linguistic conditions. All models, datasets, and code are released for full reproducibility.

Dataset Construction and Filtering

Raon-OpenTTS-Pool

Raon-OpenTTS-Pool aggregates 615K hours of English speech (240M utterances) by integrating 10+ public datasets, including both TTS-specific and ASR-oriented resources, as well as a large-scale YouTube collection (Raon-YouTube-Commons, 335K hours). Particular care is taken in recreating high-quality speech-text pairs from long-form, noisy web sources using advanced pipelines for source separation, speaker diarization, voice activity detection, and Whisper-based transcription.

Quality Filtering: Raon-OpenTTS-Core

Recognizing the critical impact of data quality on TTS, the collection is filtered using model-based metrics:

• WER (Whisper transcription accuracy as a proxy for alignment quality)
• DNSMOS (non-intrusive speech quality measure)
• Speech Ratio (proportion of active speech via VAD)

A combined score eliminates the lowest-quality 15% of samples on these axes, yielding Raon-OpenTTS-Core (510K hours, 194M segments). Aggressive filtering removes low-quality data while maximizing the scale and diversity necessary for robust TTS generation. An ablation study confirms that this moderate filtering yields optimal downstream results.

Model Architecture and Training

The Raon-OpenTTS model series, following the DiT-based continuous TTS framework, are released at two scales: 0.3B and 1B parameters. The architecture is standardized, intentionally avoiding customizations to isolate the effect of data scaling and curation. Log-mel spectrograms and character-level text encoding are utilized, with HiFi-GAN as the vocoder. Distributed training is conducted efficiently on NVIDIA B200 hardware.

Evaluation Framework

Standard Benchmarks

Raon-OpenTTS models are evaluated on widely used open and closed-data TTS testbeds, including Seed-TTS-Eval and CV3-Eval (EN and Hard-EN). Metrics focus on intelligibility (WER), speaker similarity (SIM), and perceptual quality (DNSMOS).

Raon-OpenTTS-Eval Benchmark

Addressing the limitations of narrow prompt diversity in existing benchmarks, Raon-OpenTTS-Eval comprises 6K prompt-text pairs drawn from 12 distinct datasets, systematically covering clean, noisy, in-the-wild, and expressive speech. Both objective (WER, SIM) and subjective (CMOS, SMOS) evaluations with human raters are conducted.

Results

Objective Performance

• Seed-TTS-Eval: Raon-OpenTTS-1B achieves a WER of 1.78% (second only to proprietary-data Qwen3-TTS) and leads all open-weight models in speaker similarity (SIM: 0.749).
• CV3-Hard-EN: Raon-OpenTTS-1B ranks first on both WER (6.15%) and SIM (0.775).
• Raon-OpenTTS-Eval: Raon-OpenTTS-1B consistently outperforms open-weight, open-data baselines with an overall WER of 2.81% and SIM of 0.695, demonstrating robustness across diverse conditions.

Subjective Evaluation

Human preference results (CMOS and SMOS) confirm Raon-OpenTTS-1B’s competitiveness in naturalness and speaker similarity, often matching or surpassing open-weight and some closed-weight SOTA systems.

Data Composition and Scaling Ablation

• Training with diverse pool-derived data at matched scale yields superior generalization across Clean, Wild, and Expressive regimes compared to Emilia-only models.
• Incorporating the in-the-wild Raon-YouTube-Commons subset substantially improves performance, especially in challenging real-world conditions, although providing a mixed effect in artificially noisy settings.

Implications and Future Directions

The publication of Raon-OpenTTS constitutes the largest reproducible open-source TTS reference to date, answering key questions regarding the trade-off between data scale, diversity, and quality in open-data settings. The results demonstrate that, with sufficient data and quality control, open models can match or even exceed proprietary-data-trained systems in intelligibility and speaker similarity, especially in challenging, non-studio scenarios. This work sets a new standard for reproducibility and extensibility in TTS research.

Going forward, several research vectors are indicated:

• Multilingual Expansion: Extending open-resource TTS construction beyond English.
• Domain Adaptation and Balancing: Improved weighting and balancing strategies to optimize training with heterogeneous corpora.
• Low-Resource Data Recovery: Leveraging data correction and advanced enhancement techniques to reclaim value from otherwise discarded segments.

Conclusion

Raon-OpenTTS advances the field by providing both the infrastructure and empirical validation needed for robust, high-quality, open-data TTS systems. The release of the models, datasets, and evaluation protocols enables reproducible, extensible research and empowers further development in scalable, transparent TTS foundation models. This facilitates open innovation across both academic and applied speech synthesis.