Qwen3-ASR-1.7B

Qwen3-ASR is an open family of speech recognition models; this card covers the 1.7B version, and a 0.6B version is released alongside it. Per the model card, it recognizes 30 languages and 22 Chinese dialects, identifies the language by itself, and handles speech, singing and songs with background music. One model works both offline and in streaming mode and can transcribe long audio. Timestamps need a separate helper model, Qwen3-ForcedAligner-0.6B (up to 5 minutes of speech, 11 languages). The authors ship an inference toolkit with batch and streaming modes on vLLM. The card does not state a VRAM requirement; it recommends FlashAttention 2 to save memory. On the Open ASR Leaderboard (English short-form speech, nine test sets, checked 2026-10-20) the Transformers version of the 1.7B model has an average word error rate of 5.75%, the lowest among the open models in our list. Released 2026-01-28. The weights are public without any acceptance step under the Apache-2.0 license, which allows commercial use. The authors offer an official demo Space on Hugging Face, and the code is on GitHub. Date of check: 2026-10-20.