Qwen3-ASR is an open-source series of automatic speech recognition (ASR) models, which turn spoken audio into text, developed by the Qwen team at Alibaba Cloud. The series has two recognition models, Qwen3-ASR-1.7B and Qwen3-ASR-0.6B (with 1.7 and 0.6 billion parameters), built on the audio understanding of the Qwen3-Omni foundation model, and a forced-alignment model, Qwen3-ForcedAligner-0.6B.
The recognition models identify the language and transcribe speech in 52 languages and dialects: 30 languages and 22 Chinese dialects. They handle speech, singing and songs with background music, work in streaming (live) or offline mode with the same model, and transcribe long audio. The smaller 0.6B model trades some accuracy for speed. The forced aligner predicts timestamps, the start and end time of each word or other unit, for up to 5 minutes of speech in 11 languages. A Python package, qwen-asr, runs the models with the transformers or vLLM backend.
The weights are published on Hugging Face and ModelScope, and the repository is under the Apache-2.0 license, so the models can run locally. Alibaba Cloud also offers Qwen3-ASR through its DashScope API.