Chinese AI startup VUI Labs has launched the Luna-TTS voice model, which topped the Hugging Face TTS Arena by surpassing ElevenLabs and ranked third globally in the Artificial Analysis Speech Arena, outperforming Google. The model leverages a novel Block Diffusion architecture, achieving first-place rankings globally across four voice quality metrics and two efficiency metrics. The core team, led by Professor Qian Yanmin from Shanghai Jiao Tong University and serial entrepreneur Mei Jie, aims to build a Voice Agent ecosystem with a target market of $740 billion by 2030.

Author and source: AI World

Just now, the "Chinese version of Thinking Machines Lab" was born!

A hardcore AI startup has suddenly emerged on the global voice and interaction scene, shaking up the landscape.

It is Yu Sheng Yue Ban (VUI Labs).

On Hugging Face TTS Arena, one of the most authoritative rankings in the voice industry, VUI Labs’ Luna-TTS model outperformed competitors including ElevenLabs, MiniMax, Cartesia, and other leading domestic and international companies, rising to the number one spot globally!

On the Speech Arena leaderboard by the independent AI evaluation platform Artificial Analysis, it surpassed Google to rank third globally!

This outstanding Chinese hardcore AI startup has a clear and ambitious vision: starting with high-precision TTS voice models, using voice as the most natural and immediate interaction interface, and ultimately integrating multimodal capabilities to create a fully duplex Voice Agent ecosystem.

Whether building a new generation of interaction paradigms through voice as a breakthrough or driving the evolution of multimodal foundation capabilities through real-time voice feedback, it demonstrates a path remarkably similar to that of Thinking Machines Lab. The strategies of the two companies have converged unexpectedly.

Now, VUI Labs' voice model, shortly after its release, has swiftly topped global authoritative rankings, demonstrating outstanding performance and unstoppable momentum.

How did Luna-TTS defeat giants to rank #1 globally?

Is the following dialogue between Gao Yuliang and Hou Liangping from "In the Name of the People" really between real people?

If I don’t tell you the answer, I bet you won’t be able to guess it.

Similarly, in this video, it's hard to tell the difference between a real person and AI.

The English sports commentary accurately reproduces the physical details of sound, delivering an immersive audio experience.

Voiceover for a nature documentary, seamless and natural, as if a real voice actor were behind it.

It also flows seamlessly in tech narration.

Why are today's voice AIs already pleasant to listen to, yet still sound stiff and unnatural with a robotic tone when emotionally charged?

This is not because the mainstream autoregressive (AR) model is insufficiently powerful, but because its left-to-right, word-by-word text generation logic struggles to fully capture the complexity of multiple dimensions changing simultaneously in audio.

When humans speak, emotion, tone, and breathing are intricately intertwined; however, the AR architecture enforces a unidirectional sequential generation, increasing latency for long sentences and easily leading to cumulative errors where one mistake triggers a chain of further errors.

To break through this mechanism ceiling, VUI Labs has introduced its flagship solution, Luna-TTS, offering a completely new approach that fundamentally disrupts this traditional architecture.

Luna-TTS transforms the traditional approach of generating speech token by token into a speech generation system that resembles a diffusion model, while also incorporating a dedicated real-time streaming branch.

Specifically, it first compresses audio into discrete tokens suitable for language models, then uses a Diffusion LM based on Qwen3 to simultaneously predict a large number of audio tokens; for real-time dialogue, it is further adapted into Luna-TTS Realtime, which generates audio in blocks.

After the architectural overhaul, the results were remarkable.

Six metrics ranked first globally

Luna-TTS achieves exceptional naturalness and stability across all five dimensions of sounding human-like.

It is accurate, with natural and fluent pronunciation, pacing, and sentence transitions.

Its rhythm contains sufficient variation in stress and tempo, with perfect balance of emphasis and pacing.

What’s even more remarkable is its ability to shift tone depending on context—whether expressing surprise, restraint, disappointment, or humor, it delivers a completely different voice. Every sentence feels warm and alive.

Human breaths, hesitations, soft laughs, sighs, and subtle emotional nuances in speech give real communication its texture—and Luna-TTS captures all of this in its voice.

Even with long audio transcripts, it still maintains stable expression.

Moreover, its audio and video cloning capabilities are top-tier—just a few minutes of sample audio can replicate a speaker’s voice, tone, and even breathing patterns with astonishing realism.

For example, the classic Italian cannon scene from "The Sword Drawn."

And the famous scene where Zhuge Liang scolds Wang Lang face-to-face on the battlefield.

The Chinese dub of this Godfather version is absolutely brilliant—it perfectly captures the original voiceover’s tone, conveying both grit and intensity.

The value of six firsts: four speech quality metrics plus two model efficiency metrics

In the latest technical report, Luna-TTS not only achieved first place in all four speech quality metrics on the Seed-TTS-Eval benchmark, significantly outperforming ByteDance’s Seed, MiniMax, Zhipu, and the Qwen series, but also demonstrated remarkable error correction and noise resistance capabilities in the CV3-Eval benchmark under real-world, complex environmental conditions.

These four "firsts" are extremely valuable.

In addition to performance metrics for voice quality, it also leads in two efficiency metrics: first-packet latency and real-time rate.

In summary, it offers the best voice quality, highest generation efficiency, and lowest first-packet latency.

What makes it so impressive? The key lies in the following breakthroughs.

Architecture Evolution: A Three-Step Journey from Pre-trained AR Language Models to Block Diffusion

In the TTS field, initializing from a foundational language model is an industry-recognized effective approach. Large text models inherently possess strong capabilities in multilingual text understanding, word alignment, spelling rules, and modeling complex writing systems such as those used in Japanese and Korean.

Therefore, instead of training the acoustic model from scratch, the team started with Qwen3-0.6B and designed a clear, progressive refinement roadmap.

In benchmarks such as CV3-Eval, there are numerous "difficult texts" containing long sentences, irregular punctuation, and complex colloquialisms.

For this challenging text, a bidirectional diffusion architecture is more suitable.

The reason is that the AR model has significant limitations.

Autoregressive models use left-to-right next-token prediction; once a small deviation occurs in early sampling, errors become permanently frozen in the prefix and accumulate in a snowballing manner, leading to common TTS issues such as skipped words, repetitions, and mumbled speech.

This highlights the advantage of bidirectional diffusion: unlike free-form text generation, the semantic content of TTS output is inherently tightly constrained by the input text.

Bidirectional masked diffusion enables the model to observe both past and future context globally during the denoising process, dynamically correcting local errors through multi-step confidence refinement, thereby significantly reducing exposure bias.

Tokenizer Innovation: Luna-Codec’s "Semantic Anchoring" and Acoustic Decoupling

Second, the innovation of the tokenizer.

In a multi-codebook Residual Vector Quantization neural codec, if trained end-to-end for reconstruction, the initial codebooks often prematurely capture fine spectral details, leading to highly unstable predictions from text to audio tokens.

To achieve this, Luna-Codec employs an 8-codebook architecture, reaching a frame rate of 25 Hz (200 tokens per second) at a 24 kHz sampling rate.

During tokenizer training, the team forcibly "anchored" linguistic and semantic information into the first codebook (CB1) by introducing an additional pre-trained WavLM semantic distillation loss.

CB1 is responsible for providing a stable linguistic/phonic framework, while the remaining CB2–CB8 codebooks freely capture timbre, environmental acoustics, emotional nuance, and prosodic details.

This hierarchical design from semantics to acoustics significantly reduces the difficulty of diffusion models in predicting complete audio grids from text.

Reinforcement Learning Breakthrough: GRPO Transfer on Discrete Masked Diffusion Models

One of the core algorithmic innovations of Luna-TTS is the successful adaptation of GRPO-based reinforcement learning fine-tuning to non-autoregressive discrete masked diffusion models.

In traditional autoregressive speech models, policy gradient updates can naturally proceed along the left-to-right chain decomposition of probabilities.

However, as a non-autoregressive masked diffusion model, Luna-TTS generates audio grids through multi-step iterative denoising and cannot directly compute output probabilities using the chain rule.

To solve this challenge, Luna-TTS successfully migrated GRPO to a discrete masked diffusion model, with its core lying in trajectory-aware and lexicographic reward mechanisms.

Block Diffusion and Engineering Implementation

To meet the requirements of full-duplex agents and real-time interaction scenarios, the team developed the Luna-TTS Realtime real-time speech synthesis solution.

During this process, the audio is not generated frame by frame, but rather split into segments of 1.28 seconds (32 frames). Segments are generated autoregressively (supporting KV Cache acceleration), while within each segment, 32 frames and 8 codebooks are generated in parallel through 8 to 16 denoising steps, achieving extremely high efficiency.

Its real-world performance is truly impressive: when deploying parallel CFG on two H20 GPUs, the time to first token (TTFT) is just 41.6 milliseconds, enabling the generation and decoding of the first 1.28-second audio segment with virtually no perceptible delay.

Real-time factor (RTF) as low as 0.024, meaning just 24 milliseconds to generate one second of speech—over 40 times faster than real-time—fully meeting the demands for smooth conversation.

To enable commercial scenarios such as long-text reading and full-duplex dialogue to function smoothly, the team implemented several key engineering enhancements.

For example, they implemented inference synchronization and GPU memory optimization using the vLLM-Omni framework, enabling memory decoupling and streaming chunk delivery to ensure stable responses and prevent GPU memory overflow during long-duration inference.

In addition, the team implemented text normalization: they developed a comprehensive frontend preprocessing module to handle special characters, numbers, and other text elements that models might misinterpret, ensuring that the model can always read and understand inputs accurately and smoothly, no matter how complex they are.

Data Engineering and Real Emotion Control

Luna-TTS surpasses ElevenLabs and others in conversational naturalness and expressiveness thanks to massive-scale data engineering involving millions of hours and specialized annealing fine-tuning.

First, before training the model to express emotion, the team built a high-quality multilingual speech corpus of up to 1 million hours, covering Chinese (43.4%), English (43.1%), and Japanese and Korean (combined 13.6%).

Second, the team selected approximately 100,000 hours of "ultra-high-quality" speech data for annealing training, significantly enhancing the fidelity, robustness, and natural prosody of the voice.

Finally, in the third phase of training, the team introduced expressive speech data with fine-grained annotations.

These annotations go beyond simple utterance-level emotion labels (such as [happy] happy, [angry] angry, [sad] sad) and innovatively incorporate in-line non-verbal vocalization tags (NVV), such as [laughs] (laughing), [sighs] (sighing), [gasps] (gasping), and [coughs] (coughing).

These control labels are not added as separate modules; instead, they are directly input as text prompts to the model, requiring no additional style encoder or control head.

This allows the model to naturally incorporate breathing, sighing, and emotional nuances at the right points in the text, making it sound more human and less like a robotic reading.

Thus, Luna-TTS has successfully achieved a unified balance of high-quality expressiveness and industrial-grade ultra-low latency in the TTS field.

It combines emotional expressiveness with the low-latency requirements of real-time scenarios, pushing diffusion-based voice generation to new heights.

Left hand: top scientists; right hand: seasoned business strategists—a dream-building powerhouse team

On this crowded赛道 of large AI models, while computing power and parameters are important, it ultimately comes down to people.

VUI Labs was able to develop world-leading technology in a very short time, thanks to its stellar core team.

If one word were to describe VUI Labs' DNA, it would be: dual-core driven.

One of its cores is absolute world-class scientific research capability.

Professor Qian Yanmin, founder and chairman of VUI Labs, is a distinguished professor at the School of Computer Science and the School of Artificial Intelligence, Shanghai Jiao Tong University.

This outstanding scientist, selected as a Changjiang Scholar by the Ministry of Education at just 38 years old, boasts an exceptional resume: Ph.D. from the Department of Electronic Engineering at Tsinghua University, postdoctoral researcher at the Machine Intelligence Laboratory at the University of Cambridge; recipient of the National Science Fund for Excellent Young Scholars; first recipient of the Jiao Tong Ruiyuan Young Scientist Award in Information and Space. He has long been dedicated to research and industrial application of speech dialogue and multimodal interaction technologies.

As the primary contributor, he was awarded the Wu Wenjun Award for Natural Science in Artificial Intelligence, the highest honor in China’s AI community.

In the niche and highly specialized field of end-to-end speech models, Professor Qian’s research team has ranked third globally and first nationally in paper citations over the past five years.

He has published over 350 papers, with more than 20,000 citations on Google Scholar. In the international speech technology community, he is recognized as a pioneer and serves as Chair of the Awards Committee of the IEEE SLTC Speech and Language Technology Committee.

The other core strength is its remarkable commercial monetization capability.

Scientific and technological achievements shouldn't remain stuck on lab display shelves. Mei Jie, founder and CEO of VUI Labs, is a serial successful entrepreneur with exceptional business acumen.

During his tenure as COO of Aircourse from 2017 to 2022, he grew the company’s annual revenue to over 500 million yuan in just three years, managing a large team of 800 people.

His greatest strength lies in transforming cutting-edge technologies into deliverable industry products, bridging the "last mile" from POC validation to large-scale deployment.

In addition to these two key figures, VUI Labs brings together a team of "genius minds." The entire R&D team consists of nearly 50 members, with PhD holders making up as much as 50%.

This all-around team, proficient in cutting-edge algorithms, engineering implementation, and business闭环, has laid the strongest foundation for building Thinking Machines Lab.

$740 billion market: Let AI truly earn money at work

VUI Labs has targeted real-world business scenarios since day one.

To this end, they have built a clear three-tier product system:

• Model API: Provides foundational capabilities such as speech generation, voice cloning, speech recognition and transcription, and real-time translation, with performance at the global leading level and stable, reliable service;
• Creator and Developer Platform: Provides tools for voiceover creation, long-text synthesis, multi-character performance, voice management, and project export, integrated with image, video, and music generation capabilities to cover the entire creative workflow from concept to final production;
• Voice Agent Platform: Integrates speech recognition, language models, tool invocation, and speech synthesis to provide end-to-end services—from scenario analysis and knowledge base configuration to low-latency dialogue orchestration and deployment—enabling businesses to quickly build voice agents that can converse and execute tasks.

This practical approach has been scaled and successfully implemented in scenarios such as logistics scheduling, internet insurance, and hospitality SaaS, with several customers collectively completing over 1 million real business conversations.

Luna-TTS is a true "super employee" in these scenarios—serving as a customer service representative, sales agent, meeting assistant, providing interpretation-level translation, and acting as a cabin assistant.

Estimated to reach a market size of $266 billion to $740 billion by 2030, driven by the combined capabilities of voice, vision, and collaboration.

VUI Labs is steadily positioned at this entrance.

China's next AI frontier: becoming humanity's "AI teammate"

In the future, we don’t need more cold, robotic question-answer systems—we need AI voices that breathe, think, and collaborate as naturally as humans.

Yusheng Yueban is using their state-of-the-art technology to show the world that Chinese teams are leading the global AI voice industry.

China's version of Thinking Machines has arrived. It is quietly taking up roles in hidden corners across all industries.

This time, the most natural form of interaction will become the most powerful productivity force in the business world.