Qwen3-ASR (extract text from audio)
Qwen3-ASR is an open-source speech recognition (speech-to-text) model from the Qwen team, built on the audio understanding of Qwen3-Omni. It converts speech, singing or a song into text with punctuation and timestamps. MVSep uses the largest version of the model (Qwen3-ASR-1.7B) together with Qwen3-ForcedAligner-0.6B, which aligns every recognized word to the audio.
The model recognizes 30 languages and 22 Chinese dialects and detects the language automatically. Unlike most ASR models, it was trained to handle singing and songs with background music, so it works well for transcribing lyrics. The result is the full text and a list of segments (sentences or phrases) with start and end times.
Options
- Extract vocals first - separate vocals from music before recognition. Recommended for songs with loud instruments. For speech recordings it is usually not needed. If word timestamps can't be found on the original mix, vocals are extracted automatically.
- Language - leave "auto" to detect the language automatically, or choose it explicitly if the recording contains several languages or the detection is wrong.
Supported languages
English, Chinese, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Romanian, Hungarian, Macedonian, plus 22 Chinese dialects.
Timestamps are officially supported for 11 languages: English, Chinese, Cantonese, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. For other languages, timestamps may be less accurate.
Quality
Word error rate (WER, %, lower is better) reported by the authors:
| Dataset | Whisper-large-v3 | Qwen3-ASR-1.7B |
|---|---|---|
| LibriSpeech (clean / other) | 1.51 / 3.97 | 1.63 / 3.38 |
| Fleurs (English) | 4.08 | 3.35 |
| Common Voice (English) | 9.90 | 7.39 |
| Fleurs (12 languages) | 5.27 | 4.90 |
| MLS (8 languages) | 8.62 | 8.55 |
| M4Singer (singing) | 13.58 | 5.98 |
| Popcs (singing) | 13.77 | 8.52 |
Language identification accuracy: 97.9% (Whisper-large-v3: 94.1%).
Links
- Original page: github.com/QwenLM/Qwen3-ASR
- Model weights: huggingface.co/Qwen/Qwen3-ASR-1.7B, huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B
- Paper: Qwen3-ASR Technical Report (2026)
© Qwen Team, Alibaba Cloud. License: Apache 2.0
