Alibaba's Qwen open-sources Qwen3-ASR speech models

Alibaba's Qwen team announced the open release of three models: Qwen3-ASR-1.7B, Qwen3-ASR-0.6B and Qwen3-ForcedAligner-0.6B. The weights ship under the Apache 2.0 license, together with an inference and fine-tuning framework that supports vLLM-based batch inference, asynchronous serving, streaming and timestamp prediction.

Both ASR models are built on the Qwen3-Omni foundation model. They handle language identification and transcription for 30 languages and 22 Chinese dialects, as well as English accents from multiple regions. One model runs both streaming and offline inference, and it can transcribe a single recording of up to 20 minutes. The models also work on singing and on songs with background music. Qwen reports average word error rates of 13.91% for Chinese songs and 14.60% for English songs.

Qwen says the 1.7B model is state of the art among open-source ASR models. On an internal English test set covering accents from 16 countries, the team says it outperformed GPT-4o Transcribe, the Gemini series, Doubao ASR and Whisper-large-v3. Qwen also claims it leads on Mandarin, Cantonese and the 22 dialects. Many of these comparisons rely on internal benchmarks that have not been published.

The 0.6B model is pitched on efficiency. Qwen reports an average time-to-first-token of 92 ms and throughput of 2,000 seconds of audio per second in online async mode at a concurrency of 128.

Qwen3-ForcedAligner-0.6B is a non-autoregressive, LLM-based model that predicts timestamps. It covers 11 languages and handles up to five minutes of speech. Qwen says its timestamps are more accurate than those of NeMo Forced Aligner, WhisperX and Monotonic-Aligner.