You can deploy this model and start earning money today!

KRAFTON/
Raon-OpenTTS-1B

Name: Raon-OpenTTS-1B
Availability: InStock

diffusion-transformer

KRAFTON/Raon-OpenTTS-Pool

2605.20830

cc-by-nc-4.0

Views

No views yet

Raon-OpenTTS-1B

Homepage

GitHub

Technical Report | Code | Dataset | Raon-OpenTTS-0.3B

Raon-OpenTTS is an open-data, open-weight zero-shot TTS system that performs on par with state-of-the-art closed-data models. This is the 1B variant.

Listening Demo

Zero-shot voice cloning samples from Raon-OpenTTS-1B across challenging conditions. Each row shows the generated audio, the input text, and the reference (prompt) audio.

Background Noise

Generated	Text	Reference
	That is exactly what happened
	relative to how much i think of other vacations and
	I hope not, Dick; I hope not; but smugglers don't stand at anything sometimes".

Emotion

Generated	Text	Reference
	To my surprise he began to show actual enthusiasm in my favor.

Speaking Style

Generated	Text	Reference
	Keep an eye on him.

Accent

Generated	Text	Reference
	MacDougall tapped his forehead suspiciously with a stubby forefinger.
	it occurred to me that there would have to be an accounting

Key Features

Fully Open: Both model weights and training data (615K hours, 11 English speech datasets) are publicly available for reproducible TTS research.
Competitive with Closed-Data SOTA: Ranks 1st or 2nd in WER and SIM among recent zero-shot TTS models on Seed-TTS-Eval and CV3-Eval, matching systems trained on millions of hours of proprietary data.
Robust Across Acoustic Conditions: Achieves the best average WER and SIM on Raon-OpenTTS-Eval across Clean, Noisy, Wild, and Expressive regimes.
Large-Scale Curated Data: Trained on Raon-OpenTTS-Core (510K hours), quality-filtered from Raon-OpenTTS-Pool using combined DNSMOS, WER, and VAD rank-based filtering.
DiT Architecture: Based on F5-TTS Diffusion Transformer with flow matching, enabling efficient zero-shot speech synthesis.

Model Details


Parameters	1048M
Architecture	DiT (Diffusion Transformer), based on F5-TTS
Config	dim=1408, depth=28, heads=24, ff_mult=4, text_dim=512, conv_layers=4
Training Data	Raon-OpenTTS-Core (510.1K hours)
Steps	520K updates
Hardware	48 NVIDIA B200 GPUs
Batch Size	2,688K frames (14K/GPU x 192 GPUs)
Optimizer	AdamW, peak LR 1e-4, 50K warmup, linear decay, grad norm 1.0
Audio	80-ch mel-spectrogram, 16kHz, hop=256
Vocoder	HiFi-GAN (speechbrain/tts-hifigan-libritts-16kHz)

Benchmark Results

Bold marks the best result and the Raon-OpenTTS rows. All numbers are from the technical report.

Seed-TTS-Eval

WER measured via Whisper-large-v3; SIM via WavLM-large.

Model	Params	WER (%) ↓	SIM ↑
Human	-	2.14	0.734
Seed-TTS	-	2.25	0.762
CosyVoice 3	1.5B	2.21	0.720
Index-TTS 2	1.5B	2.18	0.709
Llasa	8B	3.63	0.581
VoxCPM	0.5B	1.98	0.730
CosyVoice 2	0.5B	2.61	0.659
CosyVoice 3	0.5B	2.50	0.698
Qwen3-TTS	1.7B	1.46	0.715
Voxtral TTS	4B	2.19	0.663
MaskGCT	0.6B	2.57	0.713
F5-TTS	0.3B	2.04	0.671
Raon-OpenTTS-0.3B	0.3B	1.95	0.687
Raon-OpenTTS-1B	1.0B	1.78	0.749

CV3-Eval

WER on CV3-EN and CV3-Hard-EN; SIM via ERes2Net, DNSMOS for perceptual quality (CV3-Hard-EN).

Model	CV3-EN WER (%) ↓	CV3-Hard-EN WER (%) ↓	CV3-Hard-EN SIM ↑	CV3-Hard-EN DNSMOS ↑
F5-TTS	8.54	-	-	-
MaskGCT	7.73	41.09	0.624	3.48
CosyVoice 2	6.27	10.28	0.710	3.95
CosyVoice 3	4.96	10.77	0.740	3.98
VoxCPM	5.24	6.44	0.670	3.78
Qwen3-TTS	4.52	7.89	0.666	3.87
Raon-OpenTTS-0.3B	4.62	7.31	0.730	3.77
Raon-OpenTTS-1B	3.92	6.15	0.775	3.85

Raon-OpenTTS-Eval

4 acoustic regimes (Clean, Noisy, Wild, Expressive), 12 datasets, 6K prompt-text pairs. Overall is computed over all evaluation samples.

Model	Clean WER ↓	Clean SIM ↑	Noisy WER ↓	Noisy SIM ↑	Wild WER ↓	Wild SIM ↑	Expr. WER ↓	Expr. SIM ↑	Overall WER ↓	Overall SIM ↑
F5-TTS	2.17	0.613	3.82	0.640	136.03	0.324	3.46	0.503	25.08	0.542
MaskGCT	3.39	0.672	5.56	0.727	28.00	0.581	6.44	0.546	8.61	0.635
CosyVoice 2	2.59	0.642	4.39	0.675	49.73	0.535	3.66	0.536	11.02	0.603
CosyVoice 3	2.53	0.678	3.69	0.720	8.31	0.618	5.49	0.567	4.43	0.647
VoxCPM	2.24	0.686	3.42	0.738	43.83	0.553	2.66	0.565	9.48	0.642
Qwen3-TTS	3.38	0.684	4.60	0.726	79.14	0.528	5.81	0.527	17.59	0.626
Raon-OpenTTS-0.3B	1.57	0.645	4.03	0.700	5.83	0.571	2.53	0.570	2.93	0.623
Raon-OpenTTS-1B	1.44	0.718	3.51	0.769	5.61	0.656	2.77	0.633	2.81	0.695

Inference

For inference code and usage instructions, see krafton-ai/Raon-OpenTTS.

Training Details

Raon-OpenTTS-1B was trained for 520K update steps on 48 NVIDIA B200 GPUs using the Raon-OpenTTS-Core dataset (510.1K hours of English speech). The model uses AdamW optimization with a peak learning rate of 1e-4, 50K warmup steps, and linear decay. Gradient norm is clipped at 1.0. Waveform synthesis uses a HiFi-GAN vocoder pretrained on LibriTTS at 16kHz.

Citation

bibtex

@article{kim2026raonopentts,
  title     = {Raon-OpenTTS: Open Models and Data for Robust Text-to-Speech},
  author    = {Kim, Semin and Chung, Seungjun and Moon, Taehong and Lee, Sangheon and Ahn, Minyoung and Lee, Keon and Kim, Nam Soo and Cho, Jaewoong and Schmidt, Ludwig and Lee, Kangwook and Park, Dongmin},
  journal   = {arXiv preprint arXiv:2605.20830},
  year      = {2026},
  url       = {https://arxiv.org/abs/2605.20830}
}

License

This repository is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License.

KRAFTON/Raon-OpenTTS-1B