Qwen3-ASR Technical Report

Xian Shi, Xiong Wang, Zhifang Guo, Yongqi Wang, Pei Zhang, Xinyu Zhang, Zishan Guo, Hongkun Hao, Yu Xi, Baosong Yang, Jin Xu, Jingren Zhou, Junyang Lin

Abstract

In this report, we introduce the Qwen3-ASR family, which includes two powerful all-in-one speech recognition models and a novel non-autoregressive speech forced alignment model. Qwen3-ASR-1.7B and Qwen3-ASR-0.6B are ASR models that support language identification and ASR for 52 languages and dialects. Both leverage large-scale speech training data and the strong audio understanding ability of their foundation model, Qwen3-Omni. We conduct comprehensive internal evaluation alongside open-sourced benchmarks, as ASR models might differ little on public scores but exhibit significant quality differences in real-world scenarios. Experiments reveal that the 1.7B version achieves SOTA performance among open-sourced ASR models and is competitive with the strongest proprietary APIs, while the 0.6B version offers the best accuracy-efficiency trade-off. Qwen3-ASR-0.6B can achieve an average TTFT as low as 92ms and transcribe 2,000 seconds of speech in 1 second at a concurrency of 128. Qwen3-ForcedAligner-0.6B is an LLM-based NAR timestamp predictor able to align text-speech pairs in 11 languages, outperforming existing forced alignment models in efficiency and versatility. These models are released under the Apache 2.0 license.

Introduction

The Qwen3-ASR technical report introduces a new family of "all-in-one" speech models—Qwen3-ASR-1.7B, 0.6B, and a novel Forced Aligner—built on the Large Audio-Language Model (LALM) paradigm. These models achieve state-of-the-art open-source performance across 52 languages and dialects, rivaling proprietary APIs while offering massive efficiency gains.

Overview of Qwen3-ASR capabilities including robustness, timestamp prediction, and multilingual support.

The Evolution of Speech Recognition

Traditional Automatic Speech Recognition (ASR) has transitioned from modular systems to end-to-end (E2E) neural networks. However, even E2E models often struggle with real-world complexities such as heavy accents, noisy environments, singing voices, and the need for world knowledge to correctly transcribe named entities.

The Qwen team addresses these gaps by moving toward the LALM paradigm. In this framework, a Large Language Model (LLM) acts as the "brain," leveraging its inherent linguistic reasoning and world knowledge to interpret audio signals more effectively than traditional acoustic pattern matching.

Architecture and Training Methodology

The Qwen3-ASR architecture follows a sophisticated four-stage training pipeline designed to bridge the gap between raw audio and linguistic understanding.

The Qwen3-ASR architecture, showcasing the integration of the AuT encoder with the Qwen3 language model.
  1. AuT Pretraining: The process begins with training an "Audio-to-Token" (AuT) encoder using 40 million hours of pseudo-labeled data. This encoder utilizes an attention-encoder-decoder framework to downsample audio features into a format compatible with the LLM.
  2. Omni Pretraining: The models are initialized from Qwen3-Omni, which was pretrained on a massive 3 trillion tokens of multimodal data (text, vision, and audio).
  3. Supervised Finetuning (SFT): The model is refined to handle specific ASR output styles, including "no speech" detection and the use of context tokens for improved recognition accuracy.
  4. Reinforcement Learning (RL): Using Group Sequence Policy Optimization (GSPO), the team fine-tuned the models to enhance noise robustness and transcription stability, particularly in challenging acoustic environments.

Non-Autoregressive Forced Alignment

For the Qwen3-ForcedAligner-0.6B, the team departed from the standard autoregressive approach. Instead, they implemented a Non-Autoregressive (NAR) slot-filling formulation. By inserting special [time] tokens into the transcript, the model predicts discrete timestamp indices for every word or character simultaneously, significantly increasing speed and accuracy compared to traditional sequential tools.

Datasets and Evaluation

The team utilized a mix of massive-scale training data and diverse evaluation benchmarks to ensure robustness:

  • Training Data: 40 million hours of audio (primarily Chinese and English) and 3 trillion tokens of multimodal data.
  • Public Benchmarks: Standard sets including LibriSpeech, GigaSpeech, CommonVoice, Fleurs, MLS, and WenetSpeech.
  • Internal Stress-Tests: To move beyond the limitations of public benchmarks, the team created internal sets for 22 Chinese dialects, 16 accented English groups, and difficult scenarios like "ExtremeNoise" and "Singing Voice."

Key Results and Insights

  • State-of-the-Art Accuracy: Qwen3-ASR-1.7B achieves SOTA results among open-source models and remains competitive with leading proprietary APIs like GPT-4o-Transcribe and Gemini-2.5-Pro.
  • High Efficiency: The 0.6B version is an "efficiency beast," achieving a Time-to-First-Token (TTFT) of just 92ms. At a concurrency of 128, it can process 2,000 seconds of audio per second.
  • Precision Alignment: The Forced Aligner reduced timestamp prediction shifts by 67% to 77% compared to existing methods like MFA and WhisperX.
  • RL for Stability: Reinforcement Learning (GSPO) proved essential for reducing "hallucinations" in noisy environments where standard models might misinterpret background noise as speech.

Limitations and Future Work

While the models show exceptional performance, some limitations remain. Language Identification (LID) occasionally confuses highly similar languages like Malay and Indonesian. Additionally, the Forced Aligner currently supports 11 languages, compared to the 52 supported by the ASR models. Future work will focus on expanding language coverage and improving robustness for long-duration, music-heavy tracks.

The Qwen team has released these models under the Apache 2.0 license, providing a powerful, open-source alternative for high-performance speech applications.

Summary updated

Abstract

In this report, we introduce Qwen3-ASR family, which includes two powerful all-in-one speech recognition models and a novel non-autoregressive speech forced alignment model. Qwen3-ASR-1.7B and Qwen3-ASR-0.6B are ASR models that support language identification and ASR for 52 languages and dialects. Both of them leverage large-scale speech training data and the strong audio understanding ability of their foundation model Qwen3-Omni. We conduct comprehensive internal evaluation besides the open-sourced benchmarks as ASR models might differ little on open-sourced benchmark scores but exhibit significant quality differences in real-world scenarios. The experiments reveal that the 1.7B version achieves SOTA performance among open-sourced ASR models and is competitive with the strongest proprietary APIs while the 0.6B version offers the best accuracy-efficiency trade-off. Qwen3-ASR-0.6B can achieve an average TTFT as low as 92ms and transcribe 2000 seconds speech in 1 second at a concurrency of 128. Qwen3-ForcedAligner-0.6B is an LLM based NAR timestamp predictor that is able to align text-speech pairs in 11 languages. Timestamp accuracy experiments show that the proposed model outperforms the three strongest force alignment models and takes more advantages in efficiency and versatility. To further accelerate the community research of ASR and audio understanding, we release these models under the Apache 2.0 license.
BibTeX
@misc{shi2026qwen,
  author = {Xian Shi and Xiong Wang and Zhifang Guo and Yongqi Wang and Pei Zhang and Xinyu Zhang and Zishan Guo and Hongkun Hao and Yu Xi and Baosong Yang and Jin Xu and Jingren Zhou and Junyang Lin},
  title = {{Qwen3-ASR Technical Report}},
  year = {2026},
  eprint = {2601.21337},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.21337},
}
Figure