Luna-TTS Family: Non-autoregressive diffusion TTS beats Seed-TTS-Eval

Modern text-to-speech relies on autoregressive codec language models, which suffer from latency, error accumulation, and artificial generation order over the residual vector quantization grid. The Luna-TTS Family, detailed in a new technical report from researchers including Feng Yin and Mengxiao Bi, replaces this with a diffusion-language-model approach. Pretrained on 1 million hours of speech across Chinese, English, Japanese, and Korean, the family builds on a shared 0.6B backbone lineage and progressively adapts a pretrained autoregressive text LLM from causal to bidirectional and finally block-causal attention. The result is two variants: Luna-TTS, which is fully non-autoregressive and generates the entire RVQ token grid in a fixed number of parallel refinement steps, and Luna-TTS Realtime, which processes blocks of 32 codec frames (1.28 seconds) autoregressively while denoising each block in parallel.

Luna-TTS natively supports zero-shot voice cloning and speech editing as infilling tasks, while an annealed fine-tuning stage adds explicit control over emotion and non-verbal vocalizations. A reinforcement-learning stage applies GRPO with policy ratios over the realized denoising trajectory. On Seed-TTS-Eval, Luna-TTS achieves the best results across all four metrics among open-source and commercial systems, posting 0.73 CER / 79.7 SIM on Chinese and 1.49 WER / 76.8 SIM on English. On the harder in-the-wild CV3-Eval, it records the lowest Mandarin and English error rates. Luna-TTS Realtime delivers 0.0240 end-to-end RTF and 41.6 ms first-block latency with KV-cached blockwise generation, plus incremental audio delivery. The system also leads on most objective and human-rated metrics for NVV and emotion control against commercial rivals.