China’s Luna‑TTS Tops Global Rankings, Outperforming Google in Voice AI

VUI Labs announced the launch of Luna‑TTS, a Chinese voice‑generation model that has taken the number‑one position on the Hugging Face TTS Arena and ranked third globally on the Artificial Analysis Speech Arena, overtaking Google, ElevenLabs, MiniMax, and other leading systems.

The model’s superiority stems from a three‑step architectural evolution: it replaces the traditional left‑to‑right autoregressive token generation with a block diffusion approach that first compresses audio into discrete tokens, then predicts large batches of tokens in parallel using a Qwen‑3‑derived diffusion language model, and finally adds a real‑time streaming branch that generates 1.28‑second audio blocks with low latency.

Key technical innovations include the Luna‑Codec tokenizer, which employs an eight‑codebook residual vector quantizer operating at 25 Hz (200 tokens per second) and anchors semantic information in the first codebook via a pretrained WavLM semantic distillation loss, while higher codebooks capture timbre, environment, emotion, and prosody.

Another breakthrough is the migration of GRPO‑based reinforcement learning to a non‑autoregressive discrete mask diffusion model, enabling multi‑step denoising and parallel audio‑grid generation while preserving the benefits of policy‑gradient optimization.

Benchmark results show Luna‑TTS achieving first place on four voice‑quality metrics in the Seed‑TTS‑Eval suite and leading the CV3‑Eval benchmark in noisy, long‑sentence scenarios, as well as attaining the lowest first‑packet latency (41.6 ms) and the highest real‑time factor (RTF = 0.024) on two H20 GPUs.

To support real‑time applications, the system processes audio in 32‑frame blocks, using KV‑cache‑accelerated autoregressive generation across blocks and 8–16 steps of parallel denoising within each block, delivering over 40× faster than real time.

Extensive data engineering underpins the performance: a 1‑million‑hour multilingual corpus (43 % Chinese, 43 % English, 14 % Japanese/Korean) is refined to a high‑quality 100‑k‑hour subset for fine‑tuning, and expressive control tokens (e.g., [laughs], [sighs]) are injected directly as text prompts without extra style encoders.

The VUI Labs team, led by Professor Qian Yanmin (Shanghai Jiao‑Tong University) and CEO Mei Jie, combines top‑tier research expertise with strong commercial execution, positioning Luna‑TTS as a “super employee” for voice agents, customer service, sales, and real‑time translation across industries.

With a three‑layer product stack—model APIs, creator/developer platform, and voice‑Agent platform—VUI Labs targets a projected $740 billion market for integrated voice‑visual‑collaboration AI by 2030, aiming to make natural, human‑like AI teammates a mainstream productivity tool.