Qwen3-ASR and Qwen3-ForcedAligner Release

Qwen has open-sourced the Qwen3-ASR and Qwen3-ForcedAligner model families, introducing high-performance, multilingual speech recognition and precise timestamp alignment tools. These models leverage the audio understanding capabilities of the Qwen3-Omni foundation model to provide robust transcription and alignment across dozens of languages and dialects.

Qwen3-ASR: Multilingual Speech Recognition

Qwen3-ASR consists of two all-in-one models, Qwen3-ASR-1.7B and Qwen3-ASR-0.6B, which integrate language identification and automatic speech recognition (ASR) into a single pipeline.

Key Capabilities and Language Support

Both models support 30 languages and 22 Chinese dialects, including English accents from various countries and regions. Supported languages include Chinese (zh), English (en), Cantonese (yue), Arabic (ar), German (de), French (fr), Spanish (es), Portuguese (pt), Indonesian (id), Italian (it), Korean (ko), Russian (ru), Thai (th), Vietnamese (vi), Japanese (ja), Turkish (tr), Hindi (hi), Malay (ms), Dutch (nl), Swedish (sv), Danish (da), Finnish (fi), Polish (pl), Czech (cs), Filipino (fil), Persian (fa), Greek (el), Hungarian (hu), Macedonian (mk), and Romanian (ro).

Performance and Benchmarks

Qwen3-ASR-1.7B is positioned as the high-performance variant, achieving state-of-the-art (SOTA) results among open-source ASR models and remaining competitive with proprietary commercial APIs.

  • English: Outperforms GPT-4o Transcribe, the Gemini series, Doubao ASR, and Whisper-large-v3 on an internal test set covering accents from 16 countries.
  • Multilingual: Surpasses existing open-source models on 20 major languages, achieving the best average Word Error Rate (WER).
  • Chinese and Dialects: Leads both commercial APIs and open-source models in Mandarin, Cantonese, and 22 regional dialects.
  • Robustness: Maintains low character/word error rates in challenging scenarios, such as speech from children or the elderly and environments with extremely low signal-to-noise ratios (SNR).
  • Singing Voice: Supports full-song transcription for Chinese and English with background music (BGM), achieving average WERs of 13.91% (Chinese) and 14.60% (English).

Efficiency and Throughput

Qwen3-ASR-0.6B is optimized for the accuracy-efficiency trade-off. In online asynchronous mode with a concurrency of 128, it can transcribe 2,000 seconds of speech in 1 second (equivalent to 5 hours of speech) and achieves a time-to-first-token as low as 92 ms.

Qwen3-ForcedAligner: Precise Timestamp Prediction

Qwen3-ForcedAligner-0.6B is a non-autoregressive (NAR) LLM-based timestamp predictor designed to align text-speech pairs.

Technical Specifications

  • Language Support: Supports 11 languages.
  • Capacity: Able to predict timestamps for arbitrary units within up to 5 minutes of speech.
  • Performance: Outperforms E2E-based forced-alignment models including Nemo-Forced-Aligner, WhisperX, and Monotonic-Aligner in terms of timestamp accuracy, language coverage, and supported audio length.

Deployment and Framework

To facilitate community adoption, Qwen has released a comprehensive inference-finetune framework under the Apache 2.0 license. This toolkit supports:

  • vLLM-based batch inference
  • Asynchronous serving
  • Streaming inference
  • Timestamp prediction
  • Unified offline and streaming inference using a single model
  • Transcription of single long audio files up to 20 minutes

Sources