Confucius4-R2T2 WebGPU

Confucius4-R2T2 WebGPU

Turns Chinese and English speech into text inside your browser on WebGPU, from audio files or a live microphone. It suits people who want transcripts without sending audio to a server.

Made by mrfakename on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Confucius4-R2T2 WebGPU is a static web page that turns Chinese and English speech into text. It runs the Confucius4-R2T2 speech recognition model, a fine-tune of Qwen3-ASR-1.7B, on your own graphics card. Everything happens on your side of the connection, and the page takes both audio files and a live microphone.

Who it suits: People who need a transcript of Chinese or English speech and prefer that the audio is processed on their own machine.

Industry | Who uses it | How they use it | Example
Media | Journalists and editors | Turn spoken interviews into text | A recorded interview clip
Education | Language learners | Get written text from spoken practice | A short reading in Chinese
Software | Web developers | Study a speech recognition pipeline built for WebGPU | Reading index.html and the worker script
Accessibility | Note takers | Capture speech live from a microphone | A short spoken memo

If your browser has no WebGPU, you can skip it, since the fallback path takes minutes per clip.

Where to use it:

• A web browser with WebGPU, on the Space's page at Hugging Face
• Windows, macOS or Linux, opening index.html from a downloaded copy

Use case | Who | What they get back
File transcription | Anyone with recordings | Text from an audio file, with long audio cut into pieces of 30 seconds or less
Live dictation | Speakers and note takers | Text from the microphone as you talk
Code study | Developers | A browser-side speech pipeline to read and adapt

How to set it up:

• A browser that supports WebGPU, because the fallback is extremely slow
• Python only if you want to download the files to your computer
• Space for the model weights, which download once and stay in the browser cache

1. Install the Hugging Face hub package.
2. Download the Space with the hf command.
3. Move into the downloaded folder.
4. Open index.html in a web browser.

pip install -U huggingface_hub
hf download mrfakename/confucius4-r2t2-webgpu --repo-type space --local-dir confucius4-r2t2-webgpu
cd confucius4-r2t2-webgpu

How to use it: Open the Space page, or index.html from your copy, then give it an audio file or switch to the live microphone. The first visit downloads the model weights before any text appears.

Tool | What it does
index.html | The page you open to transcribe audio
examples/ | Sample audio, including a Chinese clip and English clips

Example: You open the page in a WebGPU browser and load the Chinese example clip. The page decodes the speech and shows the words it heard as written text.

For newcomers and experts: Start with an example clip from the examples folder. The first load is slow because the model weights download once, and after that your browser keeps them.

The audio encoder and the decoder both run on WebGPU through onnxruntime-web, using a quantized ONNX export, and the key-value cache stays on the GPU between steps. Greedy decoding runs in a Web Worker. For the microphone it uses a 16 second rolling window, reuses finished text as the decoding prefix and rolls back the last token. Weights arrive through parallel range requests and sit in the Cache API.

Audio comes from a file or the microphone
Long audio is split at quiet points
Encoder and decoder run on WebGPU
Written text appears on the page

Good to know:

• Without WebGPU it switches to WebAssembly on one CPU thread, which takes minutes per clip
• License: the demo code is Apache-2.0, while the model weights fall under the NetEase Youdao Model Use License in the MODEL_LICENSE file
• Cost to run: none named
• Data it reaches: model weights from Hugging Face on first load, with audio processed in your browser
• Not for: machines without WebGPU, or reading the changed model as endorsed by its original right-holder, who disclaims liability

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: static
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/mrfakename/confucius4-r2t2-webgpu
• Files: https://huggingface.co/spaces/mrfakename/confucius4-r2t2-webgpu/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.