Qwen3-ASR is an open-source automatic speech recognition model that supports multilingual speech recognition, language detection, and timestamp prediction across 52 languages and dialects.
The tool addresses the need for accurate, multilingual speech-to-text conversion. It provides two model sizes—a 1.7B parameter version and a 0.6B parameter version—both built on the Qwen3-Omni foundation model. The project also includes a non-autoregressive forced-alignment model for aligning text-speech pairs in 11 languages. The models handle not only speech recognition but also music and song recognition, with capabilities for language identification and timestamp prediction integrated into the same inference pipeline.
Developers should choose this tool if they need open-source multilingual ASR with competitive performance against commercial APIs. The 1.7B version is positioned as achieving state-of-the-art results among open-source models. The smaller 0.6B variant suits resource-constrained deployments. The project supports multiple inference backends including native Transformers with torch.compile optimization and vLLM for faster inference. It offers streaming inference capabilities, fine-tuning support, Docker deployment, and both local web UI and Gradio demo interfaces. The forced-alignment model provides an additional capability for applications requiring precise text-speech synchronization.
The project maintains active development with native Transformers support and torch.compile integration recently added. The codebase includes comprehensive documentation covering environment setup, Python package usage, quick inference, streaming inference, fine-tuning, deployment with vLLM, Docker containerization, and evaluation procedures. Multiple inference options are documented, from simple Python package usage to production deployment patterns. The project provides both Hugging Face and ModelScope hosting with accompanying demo interfaces and an API option through DashScope.
