Confucius4-R2T2

Licensing Requirements: The project code is released under the Apache License 2.0, permitting commercial use, modification, and redistribution subject to its terms. The model weights are governed by the proprietary NetEase Model Use License Agreement, which imposes conditions on commercial use, model improvement, and other permitted uses.

Project Overview

Confucius4-R2T2 is a low-latency, high-accuracy real-time speech recognition model developed by NetEase Youdao. It is designed for applications that require fast, continuous speech processing and converts spoken audio into text in real time. Its use cases include live captions, meeting transcription, voice input, speech translation, and speech-driven AI applications.

Underlying Technology

The model is built upon and fine-tuned from Alibaba's Qwen team's Qwen3-ASR model. It uses automatic speech recognition (ASR) technology and incorporates specialized data construction techniques, including stable-prefix data, forced time-alignment data, and token-level audio segmentation, to improve streaming recognition.

Key Features and Capabilities

• Low-Latency Streaming Recognition: Supports configurable audio chunks from 80 milliseconds to 2 seconds, allowing users to balance response speed and recognition quality.
• Stable Text Output: Uses a stable-prefix output mechanism that keeps committed text unchanged during subsequent recognition, reducing flicker in live captions and enabling immediate downstream processing.
• High-Quality Speech Transcription: Balances real-time responsiveness with recognition accuracy and also supports offline speech recognition.
• Optimized for Chinese and English: Optimized primarily for streaming recognition in Chinese and English, while also supporting French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish, and Arabic, with an emphasis on low latency and stable real-time text output.
• Context and Hotword Hints: Supports contextual prompts and hotwords to improve recognition of specialized terminology, names, and domain-specific vocabulary.
• Flexible Deployment: Supports inference backends such as vLLM and provides a GGUF model variant for llama.cpp, offering flexibility across different hardware and deployment environments.

Use Cases

Confucius4-R2T2 is suitable for live captions and dictation, meeting transcription, voice interaction, real-time speech translation, and natural language processing pipelines or large language model agents that require immediate access to speech transcripts.

Project Positioning: An open-source-code project with separately licensed model weights, designed for real-time speech applications that prioritize low latency, stable text output, and high-quality recognition.