Qwen3-ASR

Fifty-two languages and dialects fit inside Alibaba's compact speech recognition series: Qwen3-ASR ships Apache 2.0 licensed 0.6B and 1.7B models covering 30 languages, 22 Chinese dialects, and a spread of English accents, with automatic language identification built in. One architecture serves both offline transcription and streaming recognition, the latter running through a vLLM backend for real-time use, and a companion Qwen3-ForcedAligner-0.6B model produces word or character level timestamps across 11 languages on audio up to five minutes long. Installation is a single pip install of the qwen-asr package, with weights pulled automatically from Hugging Face or ModelScope on first load; NVIDIA GPUs with FlashAttention 2 support are recommended for speed, though the sub-2B sizes keep hardware demands modest. The combination of open licensing, dialect coverage, timestamps, and streaming puts it in direct competition with proprietary ASR APIs while remaining fully self-hostable, and it slots naturally into voice agents built on the wider Qwen stack.