The Chinese team has just climbed to the top of the double list!

  It's the first time in the history of the world.

Hugging Face has just been slaughtered by a Chinese team!

The voice and translations of the global open-source community on two golden racetracks, killing a Chinese black horse, left OpenAI, Google, and Great Britain behind.

It's easy to use.

Yes, that's what you did in the fourth and sixth grade, and you were checking the word every day.

Voice Recognition Trending, R2T2 tops the list

Translated by Trending, T3PO tops the list

Released in just five days, the open-source real-time voice recognition model R2T2 and the streamlined translation model T3PO, which has violently climbed to the top of the global polling platform Hugging Face, double-sitting on voice recognition and translating the thrones of two Trending lists!

The two lists are placed side by side, and the row is filled directly.

• Speech recognition ranked R2T2 at number one, OpenAI's Whisper-large-v3 dropped to sixth, while the British automaker Nemotron dropped to eighth.

• The translation chart, T3PO is also at the top of the list, and Google, Meta, the old leader of the translators in the world known as the Jiaxing Lake, has been squeezed to the back.

There is a big ambition behind these two models.

From saying, to listening, to translating, the voice agent is going to use the real voice, it's going to be a breath of fresh air.

This week after the summit, the overseas developer circle was boiling.

Within a few days of the R2T2 launch, polar explorers from around the world, as if discovering a new continent, rushed up to give it a "repair path".

Davidxifeng, the first man-of-the-match hardcore superintendent, submitted the code to the nearly 3,000-star C++ audio framework audio.cpp, rewriting the core mechanisms of R2T2 from scratch in C++. On September 19, the code officially joined the main branch.

R2T2 has completely wiped out Python, and even expensive graphics cards can't be used, and Apple's M3 chips can be up to eight times faster in real time.

Shortly thereafter, a series of offline tools emerged. The Core ML adaptation also produced a fully offline subtitle tool capable of translating 30 languages into English in real time, and the Mac speech input tool TypeWhisper wrote offline plugins specifically for it.

Even the official Hugging Face was shocked by the wave. They took the stage and actively built a ZeroGPU online demo for R2T2 and added free computing power.

ZeroGPU online demo on Hugging Face

Overseas quantification circuits also added a bit of work, driving out the full GGUF quantification version from Q2 to Q8, llama.cpp, Ollama can run.

This treatment was previously only available to top-of-the-line open-source models.

Charlie Wargnier, a preacher at Streamlit, wrote a long post at night, and at the beginning of the line, the voice agent had a very important rewriting bug, which they secretly altered. Now, the Chinese team opened a model and gave the question a second!

Alvaro Cintas, the AI blogger, said that only 2B parameters, 200 milliseconds of delay, and that every word that is spit out in the script is unrepentant and 100% open-source.

A voice-recognition model that makes the world's polar bears move like crazy is a hardcore reality that most of its peers haven't been trained in.

"Give Zhang Wei three hundred bucks, no, it's three thousand bucks, ready for the September rent".

It's a half-hearted thing that everyone has done.

Whisper uses a common "hypoflowing" method of repeating the whole sentence from scratch every few seconds.

On the other side of the Whisper, the subtitles flickered madly, listening to the edge change. The amount was once converted to nine thousand yuan, after more than half a second it was converted back to three thousand, if the assistant was quick, the money was turned out by nine thousand.

On the R2T2 side, the words are not moving as long as they are on the screen, the receiver, the amount, the note is a full pair, as is the Thai.

Another scene, set the bell. "Tomorrow at 6:30 a.m., he calls me up, and it's seven o'clock, and it rings again every five minutes".

Whisper jumped back and forth between 6:30 a.m., 7:30 a.m. and 7 p.m. and finally heard "7 o'clock" and "play again" and "think again". R2T2 was still in place.

Whisper's name has been changed several times between "Public Guard Pudding" and "Maj Babo Pudding".

The scenes changed one by one, whispered on each side, and R2T2 didn't change a word.

When the play opens, the first few words suddenly change, and the patience is gone.

But if the subtitle is followed by an AI Agent who transfers, buys tickets, navigates, and changes a word on the front, it's a disaster.

So in the Voice Agent era, "no regrets" became the deadliest need.

There's an old rule in the human conveyance circle that says that words that go out can never be recovered, and that translations can only be corrected with a hard-headed back and forth. This rule, which AI has not learned in the past, R2T2 has partially pinched it.

The full title of R2T2 is Real Real-Time Transcription, meaning "real, real, real-time transcription", and the name itself goes through the cracks.

Its solution is violent aesthetics, forcing the model to learn two moves, Commit (drop) and Wait (shut up).

For every 80 millisecond of audio that comes in, the model has to do a soul torture.

Listening, committing directly, submitting the words as if they were engraved on a stone and not ready to move.

The test showed that the same basic model, without doing this training, directly hardened it to the side of the word every 160 milliseconds, with a Chinese text error rate of up to 39.72%, which is basically nonsense.

After using the "longest stable forward" (LSP) method, the error rate dropped directly to 3.48%, almost ten times!

It's better to pull the other models in comparison than to reduce the impact.

Similarly, a word is uttered every 160 milliseconds, with an error rate of 20.64 on the Nemotron 3.5, 29.59 on the Mistral's Voxtral Mini, and only 6.42% on the R2T2.

Chinese test set, 160 milliseconds of word error and delay configuration

Not only that, the team also pulled in two of the leading commercial APIs to play the game.

They can be reversed at any time, with an English error rate of 6.20%, a delay of 0.95 seconds at the start. R2T2 stands for "falling unrepentant", and the error rate is hard pressed to 6.13%, with a delay of only 0.55 seconds.

quick and quick and unrepentant. If anyone turns away from his opponent's game, who can blame him?

English test set, contrast between true-flow and pseudo-flow

Accuracy and delayed Pareto front, R2T2 in the lower left corner

We went directly to the official online demo of the paper, where we were able to present the various claims of the drilling rigs to it.

In any case, the words on the screen only go backwards and backwards without ever changing. The four words "falling unrepentant" are certainly not white crying.

The Chinese pound-for-pound shopping instruction, "buy a Type-C charging line in Jindong, one meter five, don't buy a Lightning", didn't miss a word in English.

If you encounter a person's name or product name, fill in the "technical vocabulary" in the demo, and the words "Hugging Face" and "Hugging Face" are combined, you'll be able to identify them.

If R2T2 is "listening", then T3PO is "translating".

T3PO's full name is simulTaneous Translation via pareTo Policy Optimization, and along with R2T2, it is a tribute to R2-D2 and C-3PO.

The most important contradiction in the story is that it is easy to translate wrongly, and you have to wait for the correct translation.

T3PO's ultimate trick is to let the model come up with its own ideas, interpret the context well enough, wait a second and practice this "read or write" decision to the best of its ability.

T3PO quality and latency on five test sets, the red line being T3PO

R2T2 was responsible for throwing a piece of added text into the manuscript, and T3PO got the translation on the spot.

The entire link has evolved from a sequential pattern of "talking, recognizing, translating" into an extremely fast, full-fledged pipeline of "talking, listening, understanding, translating".

We used T3PO to test a circle, cut back into Chinese English, and after speaking the original language for more than a second, the translations followed, and it slipped like a real human being sitting next to it.

Then came the scariest scene in the story, and the speakers changed their minds.

"Tomorrow morning at ten o'clock, please, it's two o'clock in the afternoon, in the third floor meeting room, remember to bring your computer".

T3PO has translated "at ten o'clock in the morning" and can't get it back, it's like a real story, it takes the word "oh, sorry" and adds it at two o'clock in the afternoon.

It's hard to beat the newspaper's financial report.

"In the third quarter of this year, the company generated revenues of 3.750 million yuan, a growth of 23.6 percent, and net profits of 120 million yuan", it said.

Even "So, um, basically, like, you know" is not hard to translate into a clean, straightforward phrase: "Maybe we should postpone the release date to next week".

For the past two years, people have been saying "Chinese models are great", and that's basically the main list of big models.

This summer's report on Hugging Face showed that for most of 2026, the largest open-source model released by a Chinese lab outperformed its U.S. counterpart in parametric terms.

Today, the leadership of the Chinese model is spreading all the way from the main list to voice, translation, multi-modules, and video. This double-header is arguably one of the most profitable puzzles from "leading the main list" to "leading the entire spectrum".

And this puzzle has been in the works for a long time.

In 2023, the country's first large-scale educational model was launched, "the child". In May this year, the open-source multimodal reasoning engine, and in June, the open-source TTS engine, Confucius 4-TTS, which supports zero-sample speech cloning in 14 languages, complemented recognition and translation with R2T2 and T3PO in September.

From speaking to listening and translating, the complete closure of speech interaction took about a year and a half for hardcore students to open.

This capability has been used in real-life situations for a long time. Some AI transmissions have accumulated to serve 25 million users, doubling the number of users per year.

On September 16th, at the AI Open Day, CEO Zhou Feng said, "Models are smart, agents are capable, and workflow is reliable".

The future of AI is to meet people, communicate with them, make phone calls. In this war against the ultimate entry of Voice Agents, whoever has a pair of ears that can listen and fall unrepentantly is stuck in the throat of the age.

And these two ears, which the Chinese team has created, are completely open to the world.

The story of China's AI, also a one-man show on the main list, has flourished across the entire spectrum.

Now that R2T2 and T3PO are fully available on GitHub and Hugging Face, go for a quick online demo and get a feel for what real cyberspace is!