Transcription
Confucius4-R2T2
Local streaming transcription with NetEase Youdao's Confucius4-R2T2 model, run through an audio.cpp server on your Mac.
About
Confucius4-R2T2 is a community plugin by scriptease. It streams microphone audio to an audio.cpp audiocpp_server that runs the Confucius4-R2T2 model on Metal. The transcript is append-only: text that has appeared is not rewritten later. The model is optimized for Chinese and English.
The plugin supports live dictation and file transcription. Translation is not supported.
Requirements
The published macOS package requires TypeWhisper 1.7.0 or newer, macOS 14 or newer, and Apple Silicon. No API key is needed.
The model weights are licensed by NetEase Youdao under the Confucius4-R2T2 Model Use License. By downloading a model you accept it, so read it first. The license files are downloaded with each model.
Models
| Model | Download | Notes |
|---|---|---|
| Q4_K_M | 1.19 GB | Smallest and fastest. Community quantization by Nairod785. |
| Q8_0 | 2.48 GB | Recommended. Near full quality. |
| F16 | 4.09 GB | Full precision, largest. |
Setup
- Install Confucius4-R2T2 from Settings > Integrations > Discover and enable it.
- Keep Built-in selected and click Download next to a model. The plugin downloads audio.cpp v0.9.0 and the model from Hugging Face, then checks size and SHA-256 of both.
- Select Confucius4-R2T2 as the transcription engine in Settings or in a profile.
- Start dictating. The server starts with the first dictation, listens only on
127.0.0.1, and stops with TypeWhisper.
Own server
Choose Own server to use an audiocpp_server you run yourself, version 0.9.0 or newer. Enter the Server URL and the Model ID from your server config, then click Test Connection. The model entry must use streaming mode.
Audio is sent only to the server configured here. For a server on another machine, use https://: plain http:// sends the audio unencrypted over the network.