A robust extension for ComfyUI, the Qwen3-ASR integration provides advanced speech-to-text transcription and language identification capabilities. It is designed to deliver high-accuracy results across 52 languages and dialects, including various English accents and multiple Chinese dialects.
- Supports the Qwen3-ASR models (0.6B and 1.7B) for superior transcription accuracy.
- Offers word-level timestamps and automatic language detection, enhancing the usability of transcriptions.
- Incorporates VRAM-optimized features like FlashAttention 2 for efficient processing and reduced memory usage.
Context
This tool serves as a high-performance integration for the Qwen3-ASR model family within ComfyUI, focusing on delivering state-of-the-art automatic speech recognition (ASR) and language identification. Its primary purpose is to facilitate accurate transcription of audio inputs while providing detailed timing information for each word.
Key Features & Benefits
The extension boasts several practical features that enhance its functionality:
- High Accuracy: By utilizing the Qwen3-ASR models, users can achieve precise transcription results across a wide range of languages and dialects.
- Word-Level Timestamps: This feature, when used in conjunction with the Qwen3-ForcedAligner, allows users to receive detailed timing information for each word, which is critical for applications that require precise synchronization with audio.
- Automatic Language Detection: The ability to automatically identify the language being spoken simplifies the user experience, allowing for seamless transcriptions without the need for manual input.
Advanced Functionalities
The integration supports flexible precision settings (bf16, fp16, fp32), enabling users to optimize performance based on their hardware capabilities. Additionally, the FlashAttention 2 feature significantly reduces VRAM usage while accelerating inference, making it suitable for users with limited resources.
Practical Benefits
By incorporating this tool into their workflows, users can expect improved efficiency and control over their audio transcription processes. The ability to handle long audio segments with automatic resampling ensures that the tool can adapt to various audio qualities and lengths, enhancing overall productivity in ComfyUI.
Credits/Acknowledgments
This project is developed under the MIT License, with contributions from the original authors and the community. The Qwen3 models are governed by the Qwen License Agreement, ensuring compliance and proper usage.