Qwen3-ASR-1.7B

Qwen3-ASR-1.7B is the flagship model in the Qwen3-ASR family from Alibaba Cloud's Qwen team. It is a Large Audio-Language Model (LALM) post-trained from the Qwen3-Omni foundation model, pairing an audio encoder with a Qwen3 transformer decoder via a learned projector.

• 52 languages/dialects: 30 major languages (Chinese, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Hungarian, Macedonian, Romanian) plus 22 Chinese dialects (e.g., Cantonese HK/GD, Wu, Minnan, Sichuan, Shanghai, etc.).
• All-in-one: joint language identification + transcription.
• Robust recognition in noisy/complex acoustic conditions, singing voice, accents, and songs with background music.
• SOTA among open-source ASR models; competitive with GPT-4o-transcribe class commercial APIs.
• Unified streaming & offline inference; long-audio support up to ~20 minutes per pass.
• Companion Qwen3-ForcedAligner-0.6B provides multilingual word/sentence timestamp alignment.

Ships with qwen-asr PyPI package, vLLM backend for batch/async/streaming serving, and Docker images.

Meeting transcription, podcast/video subtitling, multilingual voice agents, call-center QA, media monitoring, and research on large audio-language models.