SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Qwen3-ASR - Alibaba Tongyi's Open-Source Series of Speech Recognition Models

Author: AI小集 · Original source: AI工具集

What is Qwen3-ASR

Qwen3-ASR is a series open-sourced by the Alibaba Cloud Tongyi Qianwen teamspeech recognitionmodels, including two ASR models, the 1.7B high-precision version and the 0.6B efficient version, as well as the dedicated Qwen3-ForcedAligner-0.6B forced alignment model. The models support recognition of 52 languages and dialects, integrated streaming and non-streaming inference, and perform robustly in complex scenarios such as strong noise, fast speech, and singing. Its 1.7B model achieves open-source SOTA in Chinese, English, and dialect recognition, while the 0.6B model supports 128 concurrent requests with 2000x throughput, processing 5 hours of audio in 10 seconds.

Qwen3-ASR

Main Features of Qwen3-ASR

  • Multilingual Recognition: Supports speech recognition and language identification for 30 major languages, enabling unified multilingual modeling.
  • Dialect Coverage: Supports recognition of 22 Chinese dialects, covering Cantonese, Wu, Minnan, and local accents from various provinces.
  • Accent Adaptation: Adapts to English accent differences across 16 countries/regions, improving cross-regional recognition capability.
  • Dual-Mode Inference: Supports integrated streaming and non-streaming inference, and can process up to 20 minutes of audio in one pass.
  • Scenario Robustness: Possesses robustness in complex scenarios, handling challenges such as strong noise, low audio quality, fast speech, and speech from the elderly and children.
  • Singing Transcription: Supports singing recognition with background music, and can transcribe entire Chinese and English songs.
  • Timestamp Alignment: Provides word-level/sentence-level timestamp alignment functionality, meeting the needs of subtitle generation and audio editing.

Technical Principles of Qwen3-ASR

  • Speech Encoding Layer: Uses an innovative pretrained AuT speech encoder to extract high-level acoustic representations, replacing traditional Fbank features and enhancing generalization to noise and accents.
  • Multimodal Foundation: Built on the Qwen3-Omni multimodal large model, using its cross-modal understanding capability to achieve direct speech-to-text mapping without the need for traditional HMM/GMM pipelines.
  • Training Paradigm: Establishes a general acoustic-semantic space through large-scale multilingual pretraining, then performs fine-grained fine-tuning for scenarios such as dialects, singing, and noise, jointly optimizing language identification and speech recognition tasks.
  • Efficient Inference: The 0.6B model adopts the vLLM acceleration engine, supports batch inference and asynchronous services, and achieves 2000x throughput under 128 concurrency; the streaming version adopts a chunked caching mechanism to balance real-time performance and accuracy.
  • Forced Alignment: ForcedAligner is based on a non-autoregressive LLM architecture and predicts timestamps through parallel decoding, achieving a single-concurrency RTF of 0.0089, with accuracy surpassing traditional CTC and WhisperX solutions.

Qwen3-ASR Project Address

  • Project Website：https://qwen.ai/blog?id=qwen3asr
  • GitHub Repository：https://github.com/QwenLM/Qwen3-ASR
  • HuggingFace Model Library：https://huggingface.co/collections/Qwen/qwen3-asr
  • Technical Paper：https://github.com/QwenLM/Qwen3-ASR/blob/main/assets/Qwen3_ASR.pdf

Application Scenarios of Qwen3-ASR

  • Smart Meetings: The model can transcribe multi-person meeting content in real time, support mixed Chinese and English as well as various dialect accents, and automatically generate meeting minutes with timestamps.
  • Video Subtitles: Generate accurate subtitles for films, TV dramas, short videos, and live streams, supporting recognition of sung content with background music and multilingual subtitle translation.
  • Telephone Customer Service: Stably recognize customer speech in high-noise, low-audio-quality call environments; the model supports real-time streaming transcription and keyword extraction.
  • Smart Speakers: Adapt to non-standard pronunciation from the elderly, children, and others, support far-field sound pickup and dialect interaction, and improve the smart home voice control experience.
  • Legal Forensics: Perform high-precision transcription of audio evidence in complex acoustic environments and provide word-level timestamps for courtroom evidence presentation and content verification.