Upload a song, get a chart and a dance: Golden Saina's rhythm game Yingwutuan ships a music-to-dance model to ordinary players
At 18:00 China time on November 19, Shanghai Golden Saina AI Technology launched Yingwutuan (鹦鹉团) on iOS and Android nationwide. It is a 3D social rhythm-and-dance game. By launch day it had climbed into the top 20 of the App Store's music chart in China, and it passed 5,000 followers on TapTap within its first day on the store.
What makes it worth covering is not the game itself but the model underneath it: a self-developed multimodal model Golden Saina calls Parrot AI (鹦鹉 AI). This is the first time that model has been opened to a mass audience, and it changes who gets to make content. A player uploads a song, the model reads its beats and section structure and builds a playable chart. The player taps once, the model pulls from a library of more than 1,000 motions and assembles a dance sequence that fits the music, and then lets the player edit it section by section.
A note on sourcing first. Launch time, version and product features can be checked directly on the App Store listing, the official site (auditionsgo.com) and TapTap. The model's capability list, data scale and technical detail come from a consistent set of Chinese media reports published between November 19 and 20. This article is compiled from multiple media reports for the technical portion; the company's own channels disclose only product features and algorithm registration numbers.
TL;DR
The short version: this is the first time a music-to-dance model has shipped in a form anyone can use, rather than as another paper or demo.
Four numbers:
- 18:00, November 19 — nationwide launch on iOS and Android; the App Store build is currently version 1.1
- 1,000+ built-in official motions, the library behind the one-tap choreography
- 500TB of proprietary music-and-dance data (images, audio and video accumulated over years of running the SNH48 GROUP), used for initial training
- 200,000+ multilingual, multi-track original songs, used to teach the model how vocals, accompaniment and drums are layered
The problem it addresses is not whether AI can produce a good-looking dance. It is a more mundane one: getting a dance from a song to a finished performance used to require a chart designer, a choreographer and an art team, and ordinary players had no way in.
Why this is a structural change for the genre
For more than two decades the rhythm-game model barely moved. A publisher licenses songs, chart designers hand-build each chart, artists design outfits, motion comes from mocap, and players compete for scores inside content that is already finished. The only thing a player decides is how well they play, not what they play.
That structure is stable but has an obvious ceiling. The song library depends on licensing budgets, the number of charts depends on how fast designers can work, and motions and outfits depend on art and mocap schedules. Content is scarce, so content always comes from the publisher.
Yingwutuan opens four layers to players:
- Songs: upload a local track and the AI generates a chart, so the library no longer depends on what was licensed
- Dances: one tap generates a motion sequence matched to the music, editable afterwards section by section
- Staging: players arrange positions, lighting and camera moves
- Appearance and outfits: regional dyeing, freehand painting across the whole garment, decals, material and weave detail, and hundreds of color schemes
The first two are genuinely new, because those tasks used to require professionals. The last two have precedents elsewhere in the market and count as experience polish.
Put differently: the pitch is not "AI plays the song for you." It is "AI turns part of a chart designer's and a choreographer's job into a single click." The difference matters — the first lowers difficulty, the second changes how content gets produced.
What Parrot AI actually does
According to the media accounts, Parrot AI is a vertical multimodal model built specifically for the music-and-dance space. Its capability set includes text-to-music, music-to-dance motion generation, text-to-dance, TTS, voice cloning and personalized speech synthesis, plus text-to-image and controllable style image generation on the visual side.
Inside the game, three of those capabilities map to three features.
Automatic charting. After a player uploads a song, the model does four things: detect beats and accents; identify section structure (intro, verse, chorus, interlude); generate chart density that matches the song's mood; and set a sensible difficulty curve. Technically this is audio understanding; as a product it is game design — charts that are too dense overwhelm the player, charts that are too sparse feel flat, and accents that do not line up with the rhythm make the controls feel wrong.
Automatic choreography. Built on the music-to-dance capability, a single tap produces a motion sequence matched to the music. The model calls on and combines motions from a library of more than 1,000 built-in clips to form a complete dance, and the result can be edited section by section. The design intent is clear: the AI produces a usable first draft, and creative control stays with the player instead of being taken away.
Automatic image generation. Used for posters, covers and shareable assets, and to help with decals and patterns on outfits. This looks like the least important of the three, but it closes the loop — people who make something need a way to show it.
One technical constraint from the reporting deserves its own mention: generated motion has to be mapped onto a virtual character, so the center of gravity, the joints and the motion trajectory all have to obey the laws of movement, and the system has to avoid the character clipping through itself. That constraint is what separates "generated" from "usable," and it is where a vertical model diverges from a general one.
Where the data comes from
This is the hardest part to copy. Per the reporting, Golden Saina's data assets come in three blocks:
- 500TB of proprietary music-and-dance data, from a long-running partnership with Star48 Media (丝芭传媒). The bulk is images, audio and video accumulated while operating the SNH48 GROUP, technically annotated and used for initial training;
- More than 200,000 multilingual original multi-track songs. Multi-track means vocals, accompaniment and drums can be separated, which lets the model understand how a song is built internally;
- Large volumes of single-person and multi-person motion capture data, the sample base behind the music-to-dance capability.
What the three have in common is that they were accumulated over years inside a real business rather than scraped from public corpora. For a field where music structure, human motion and real-time rendering all have to be right at once, that kind of data sits closer to the problem than general-purpose training material.
There is supporting infrastructure as well: facial driving for digital humans, pose simulation, parametric voice modeling, image recognition and processing, controllable style attribute decoding, and other modules. Golden Saina says it holds patents related to music-to-dance technology. The official site's footer also carries the algorithm registration: "Parrot human music-and-dance generation algorithm, CAC filing No. 310106775081801240017."
How to use it: what an ordinary player actually does
If you just want to try it, the path is short:
- Install and sign in. Search for Yingwutuan on the App Store or an Android store. It is free to download with in-app purchases, the download is about 2GB, and the age rating is 12+. There is no subscription price to evaluate before you start.
- Pick your own song. Upload a local track and wait for the AI to build a chart; difficulty is set from the track, so obscure songs get charts too.
- Open the stage builder. Tap to generate a dance; the model assembles a full sequence from the 1,000+ motion library.
- Edit. Rework the sections you do not like, or arrange positions, lighting and camera yourself.
- Publish. Generate a poster or short video asset and post it to the creative community, where other players can copy your dance and your custom chart with one tap.
Steps one to four are creation. Step five is the real lever: one-tap copying turns any single player's output into someone else's starting point.
Who it is for, and who it is not for
A good fit if you:
- play rhythm games and keep running into songs the official library will never carry;
- make short-form video and want ready-made choreography you can still edit;
- are a small team or club testing user-generated music-and-dance content on a budget.
Not a good fit if you:
- need professional-grade choreography. The model calls on and combines motions from a library of 1,000+, so the movement vocabulary is bounded by what is in that library;
- need commercial licensing. Creating inside a game and using the result commercially are separate questions, and the terms have to be read on their own;
- want a full music video generated end to end. This is an in-game creation tool, not a video generation model.
Cost and time to try
The barrier to entry is close to zero: the app is a free download with optional in-app purchases, setup takes a few minutes, and the first generated chart plus dance takes roughly as long as waiting for the AI to process the track. Editing a section afterwards is a few taps. The real cost is the time you spend iterating on choreography once you care about the result.
Pitfalls to watch
- Do not read this as a model release. What shipped is a game; the model is the foundation. There is no model card, no parameter count and no published benchmark. Readers looking for technical specifications will come away empty-handed.
- The technical detail is media-sourced only. Parameter counts, training methods and evaluation metrics are not public. Every capability description above falls under "compiled from multiple media reports," not a first-party technical document.
- Test the motion library yourself. How many genuinely different dances 1,000+ motions can produce depends on the combination logic, and that can only be judged hands-on.
- Launch-day load is normal. Some players reported login problems on the App Store on day one, which is common for a new online game and does not change the assessment here.
What to watch next
Golden Saina is also building out IP: an AI-driven virtual idol group called AOD, five units and 23 members across anime and photoreal styles, with music videos, singles, variety content and photo books produced with the same technology stack. The game serves player creation; AOD serves professional content production — two applications of one foundation.
The real signal is not launch numbers. It is two things: whether the charts generated from player-uploaded songs are genuinely playable, and whether one-tap copying turns into a content loop. If both hold, the supply logic of the rhythm-game genre really has changed. If not, this is a rhythm game with more elaborate features.