Qwen3-ASR and Qwen3-ForcedAligner Open-Sourced: Robust Streaming Multilingual Support

The Qwen3-ASR family consists of Qwen3-ASR-1.7B, Qwen3-ASR-0.6B, and Qwen3-ForcedAligner-0.6B. Both ASR models handle language identification and speech recognition within a single model, covering 30 languages, 22 Chinese dialects, and multiple national English accents.

The 1.7B version aims for top-tier performance among open ASR models, and in internal evaluations it produced results competitive with commercial APIs. The 0.6B version focuses on balancing accuracy and efficiency, and is reported to process 2,000 seconds of speech in 1 second at concurrency 128, with time-to-first-token as low as 92ms.

• Coverage: speech, singing voice, songs with BGM
• Inference mode: unified offline / streaming, handling speech up to 20 minutes long in a single model
• Framework: vLLM-based batch inference, supporting asynchronous serving, streaming inference, and timestamp prediction
• License: Apache 2.0

Qwen3-ForcedAligner-0.6B is a non-autoregressive (NAR) timestamp predictor that performs text-speech alignment on audio up to 5 minutes long. It supports 11 languages, and in evaluations it is described as showing higher timestamp accuracy and efficiency than Nemo-Forced-Aligner, WhisperX, and Monotonic-Aligner.

Performance evaluation was conducted broadly across English, multilingual, Chinese/dialects, challenging acoustic conditions, and song transcription. On the internal English test set, it is claimed to outperform GPT-4o Transcribe, Gemini, Doubao ASR, and Whisper-large-v3, and it is also presented as delivering strong results compared to open models in the multilingual and Chinese/dialect domains.

Along with the weights for all three models, a powerful inference-finetune framework was also released, with the goal of accelerating the spread of ASR and audio understanding research.