Qwen3-ASR & Qwen3-ForcedAligner is Now Open Sourced: Robust, Streaming and Multilingual!
Qwen3-ASR family 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 accents. 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 state-of-the-art performance among open-sourced ASR models and is competitive with the strongest proprietary commercial APIs while the 0.6B version offers the best accuracy–efficiency trade-off. Qwen3-ASR-0.6B can achieve an average time-to-first-token as low as 92 ms and transcribe 2,000 seconds speech in 1 second at online async mode and 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 accelerat the community research of ASR and audio understanding, we open-source the weights of the three models so as to a powerful and easy-using inference-finetune framework under the Apache 2.0 license.
Model List#
| Model | Supported Languages | Supported Dialects | Inference Mode | Audio Types |
|---|---|---|---|---|
| Qwen3-ASR-1.7B & Qwen3-ASR-0.6B | 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), Romanian (ro) | Anhui, Dongbei, Fujian, Gansu, Guizhou, Hebei, Henan, Hubei, Hunan, Jiangxi, Ningxia, Shandong, Shaanxi, Shanxi, Sichuan, Tianjin, Yunnan, Zhejiang, Cantonese (Hong Kong accent), Cantonese (Guangdong accent), Wu language, Minnan language. | Offline / Streaming | Speech, Singing Voice, Songs with BGM |
| Qwen3-ForcedAligner-0.6B | Chinese, English, Cantonese, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish | – | NAR | Speech |
Qwen3-ASR Key Features#
Main Features:
-
All-in-one: Qwen3-ASR-1.7B and Qwen3-ASR-0.6B support language identification and speech recognition for 30 languages and 22 Chinese dialects, so as to English accents from multiple countries and regions.
-
Excellent and Fast: The Qwen3-ASR family ASR models maintains high-quality and robust recognition under complex acoustic environments and challenging text patterns. Qwen3-ASR-1.7B achieves strong performance on both open-sourced and internal benchmarks. While the 0.6B version achieves accuracy-efficient trade-off, it reaches 2000 times throughput at a concurrency of 128. They both achieve streaming / offline unified inference with single model and support transcribe single long audio up to 20 minutes.
-
Novel and strong forced alignment Solution: We introduce Qwen3-ForcedAligner-0.6B, which supports timestamp prediction for arbitrary units within up to 5 minutes of speech in 11 languages. Evaluations show its timestamp accuracy surpasses E2E based forced-alignment models.
-
Comprehensive inference toolkit: In addition to open-sourcing the architectures and weights of the Qwen3-ASR series, we also release a powerful, full-featured inference framework that supports vLLM-based batch inference, asynchronous serving, streaming inference, timestamp prediction, and more.
ASR Model Performance#
We conducted a systematic evaluation of the Qwen3-ASR series across Chinese/English, multilingual settings, Chinese dialects, singing voice recognition, and challenging acoustic and linguistic scenarios. The results show that Qwen3-ASR-1.7B achieves open-source SOTA on multiple public and internal benchmarks across several dimensions. Moreover, compared with the latest ASR APIs from multiple commercial providers, it also delivers the best performance on a number of benchmarks. Specifically:
- English: In addition to achieving top performance on common public benchmarks, we evaluated on an internally built English test set covering accents from 16 countries. Overall, it consistently outperforms GPT-4o Transcribe, the Gemini series, the Doubao ASR series, and the strongest general-purpose open-source model, Whisper-large-v3.
- Multilingual: Supports up to 30 languages. On 20 major languages, Qwen3-ASR-1.7B surpasses existing open-source models across the board, achieving the best average WER.
- Chinese and dialects: On Mandarin, Cantonese, and 22 regional dialects, Qwen3-ASR-1.7B overall leads both commercial APIs and open-source models.
- Challenging acoustic/linguistic scenarios: It remains stable and produces reliable outputs under challenging conditions such as elderly/child speech, extremely low SNR, maintaining very low character/word error rates.
- Singing voice recognition: Supports full-song transcription (Chinese/English) with background music (BGM); it achieves average WERs of 13.91% (Chinese) and 14.60% (English), respectively.
Qwen3-ASR-0.6B strikes a strong balance between accuracy and efficiency: it delivers robust performance on multiple Chinese and English benchmarks, and maintains extremely low RTF and high throughput under high concurrency in both offline batch and online async inference. At concurrency 128, Qwen3-ASR-0.6b can transcribe 5 hours speech at online async mode.
FA Model Performance#
Qwen3-ForcedAligner-0.6B outperforms Nemo-Forced-Aligner, WhisperX and Monotonic-Aligner, which are three strong E2E based forced-alignment models. And it is also proved of taking great advantages on aspects of language coverage, timestamp accuracy, speech and audio length supported.
Find more detailed results in our paper.
Qwen3-ASR Demos#
Qwen3-ASR-1.7B English Demos
| Audio | ASR Results | |
|---|---|---|
| English, low speech quality | Okay, Charles. It looks like we have a problem with the radio. What happened? Yeah, someone spilled water on their machine. I uh, yeah. Charles, can you hear us? Mamma mia. | |
| English, multiple kinds of noise | My girls, my girls, my girls, my girls. Ready? Hey, babe. Hey, babe. Where are you? I'm actually crazy traffic right now. Oh, really? Yeah. It's crazy. The freeway is completely stopped. Oh, you're still coming to my parents' house, right? Um, I can't really hear you, babe. What? Mariachi band playing live music, yeah. Babe. I can't. They're really loud, and I can't hear. Babe. What? Yeah, they're being really loud right now. I'm sorry. What were you saying? You're still coming to my parents' house, right? It actually started raining like crazy. What? It's raining and thundering like crazy, man. I can't hear shit. Where are you? Out of nowhere, it's just pouring. It's pouring. It's pouring like crazy. What? Insane. I don't think I can get anywhere today. It's crazy day. Babe, that's crazy. Where are you? Oh my God! Someone just hit my car. Come on, get in the car, Gabe. Oh my God! He's getting out of his car. He's getting out of his car. What? Hey, Mari, you just hit my car! Oh, babe, this guy's crazy. What the fuck? Out, out, out, out, out! How you like that? Oh, babe, he's beating the shit out of you. Hold on a sec, babe. Oh, you know I'm gonna get my gun. Here's a gun. Oh, babe, he shot me. He shot me in the leg. He shot. Oh, babe, oh fuck! Babe, I need a driveway. I need to get out of here. I need to get out of here. I'm driving away. Where are you? Babe, this is crazy. No, I'm okay. I'll be fine. I'm okay. I'm okay. I just had to drive. Oh shit, babe, I think I'm getting pulled over now. I think I'm getting pulled over. Pull over your vehicle. Oh my God, babe, hold on a sec. Baby, talk to me. Oh shit. License and registration, sir. Yeah, of course, officer. Of course. Ah, babe, this is the worst day of my life. I just got pulled over. Oh my God. Oh my God. Officer, police. Oh my God. I think a riot's breaking out. We got a crazy riot. People are going crazy. Get out of here! It's like a lotus matter. There's like a war going on or some shit. | |
| English, rap | Sometimes I just feel like quitting. I still might. Why do I put up this fight? Why do I still write? Sometimes it's hard enough just dealing with real life. Sometimes I wanna jump a state and just kill mics. And so these people want my level of skills, like, but I'm still white. Sometimes I just hate life. Something ain't right. Hit the brake lights. Taste of the state's right. Drawing the blank line. It ain't my fault. |
Qwen3-ASR-1.7B Multiingual Demos
| Audio | ASR Results | |
|---|---|---|
| Cross-lingual: English, French, Italian, Spanish. | I'm alone, all by myself. Je suis tout seul. Sono tutto. Estoy solo. | |
| Arabic | إطلالات مكياج عيون ذهبي لسهرات صيف عشرين واحد وعشرين بأسلوب النجمات. | |
| German | Raptorium Bergbau scheint profitierter als Monroe als Reaktion auf die wirtschaftlichen Ausfälle zu sein. | |
| Spanish | Esta prenda es amplia, recomiendo elegir una talla menor a la habitual. | |
| French | Alice et moi sommes allés à Paris voyager en train au printemps, c'était très amusant. | |
| Russian | Барсук, живущий в киевском зоопарке, совершил побег из своего вольера. | |
| Japanese | 抜群の運動神経を持ち合わせ、どんな要求にも応えてきた。 |
Qwen3-ASR-1.7B Chinese Demos
| Audio | ASR Results | |
|---|---|---|
| Chinese, fast speed | 蹦出来之后，左手、右手接一个慢动作，右边再直接拉到这上面之后，直接拉到这个轮胎上，上边再接过去之后，然后上边再直接拉到这个位置了之后，右边再直接这个位置接倒过去的之后，再倒一下，然后右边再直接抓住这个上边了之后，直接从这边上边过去了之后，直接抓住这个树杈，然后这个位置直接倒到这个树杈。 | |
| Chinese, tongue twister | 广西壮族自治区爱吃红鲤鱼与绿鲤鱼与驴的出租车司机，拉着苗族土家族自治州爱喝自制的刘奶奶榴莲牛奶的骨质疏松症患者，遇见别着喇叭的哑巴，打败咬死山前四十四棵紫色柿子树的四十四只石狮子之后，碰到年年恋牛娘的牛郎，念着灰黑灰化肥发黑会挥发，走出香港官方网站设置组，到广西壮族自治区首府南宁市民族医院就医。 | |
| Chinese, heavy noise, low quality | 拨号，请再说一次，请说出您要拨打的号码。幺三五八幺八八七五七。一三五八二八八八幺八八。纠正纠正。九六九。纠正纠正，不是九六。 | |
| Chinese, song with BGM | 黑的白的红的粉的紫的绿的蓝的灰的，你的我的他的她的大的小的圆的扁的，好的坏的美的丑的新的旧的，各种款式各种花式，让我选择。飞得高喽，越远越好，天都沦陷，他就死掉，说明多能高兴就好，喜欢就好，没大不了，越变越小，越来越小，快要死掉也很骄傲。你不想说就别再说，我不想听不想再听，就把一切誓言当作气球一般随它而去，我不在意不会在意，随它而去随它而去。气球飘进眼里，飘进风里，结束生命。气球飘进爱里，飘进心里，慢慢死去。 | |
| Cantonese, noise | 今次寻寻觅觅，终于揾到my princess，肯借个场俾我哋玩。你知啦，喺香港地喺繁忙时间要揾个场嚟拍嘢系非常之难嘅。再一次多谢你哋，亦都好多谢片入边嘅每一个人。下一次我哋斗啲咩好？喺下面留言话我哋知啦。拜拜。 |
Qwen3-ForcedAligner-0.6B Demos
| Audio | ASR Results | Clipped Samples | |
|---|---|---|---|
| English, 83s | With that next, are these frozen wood frogs? And if somebody sucks on them, it'll get rid of their fever. But make sure they're frozen, because I guess when they thaw out, they're useless. He just his immediate response is like, "You're crazy," and she's like, "Yeah." And he was the one at the very beginning that he battles with the chokes against. That was proof. Dr. Palacios is a National Board certified teacher and leading expert in early childhood education. And with over five decades of experience, she is a pioneer in the field of dual language learning and specializes in curriculum planning and instructional design. Um, and he said that being the Pez outlaw is just not what he needed to do anymore. He needed to step up and be there for Kathy. Um, he needed to stop worrying about all this nonsense. And he says he pretty much just becomes a new man because he loves her so much, which just look for value in life, to look for instruction, to look for improvement, to look for development, to look for connection with other people. You you communicate and you take this on board. Because it's just like back and forth. But then they threw some shade as well, saying that Nvidia kind of like cherry picked, um, floating like a different a type of floating point that only is. | frozen:wood:frogs:Nvidia: | |
| Chinese, noise, 28s | 在他十一岁的时候，母亲因为发生交通事故撒手人寰。桂兰的母亲敲了几下杯子，铁柱的手也不自觉地动了起来。你这小子是烟瘾犯了吧？被看穿的铁柱只能坦白自己正在戒烟。父亲一瞅，给铁柱提了一个建议，嘎嘎好使。不信你可以尝试一下这个绝技，就是母亲的催眠疗法，保证让你满意。铁柱听完，觉得他们是在扯淡。这种催眠疗法我根本不需要尝试。不管咋说，几个人聊的还是其乐融融。他们非常欢迎铁柱到这里做客，但这个女佣却显得异常诡异。 | 十:一:异:常: |
ASR Results with timestamps 1
With[0.48,1.04] that[1.04,1.28] next[1.44,2.00], are[2.16,2.32] these[2.32,2.48] frozen[2.48,3.20] wood[3.36,3.68] frogs[3.68,4.32]? And[4.64,4.96] if[4.96,5.20] somebody[5.20,5.60] sucks[5.68,6.24] on[6.24,6.48] them[6.56,6.96], it'll[7.04,7.28] get[7.28,7.52] rid[7.52,7.60] of[7.60,7.76] their[7.76,7.92] fever[7.92,8.40]. ......[omitted]...... saying[72.88,73.44] that[73.44,73.60] Nvidia[75.12,75.68] kind[75.68,75.84] of[75.84,75.92] like[75.92,76.08] cherry[76.08,76.40] picked[76.40,76.80], um[77.28,77.76], floating[78.64,79.04] like[79.04,79.20] a[79.20,79.20] different[79.20,79.68] a[79.68,79.68] type[79.68,79.92] of[79.92,80.08] floating[80.08,80.40] point[80.40,80.80] that[80.80,80.96] only[80.96,81.60] is[81.92,82.32].
ASR Results with timestamps 2
在[0.00,0.16]他[0.16,0.32]十[0.32,0.48]一[0.48,0.64]岁[0.64,0.80]的[0.80,0.88]时[0.88,0.96]候[0.96,1.12]，母[1.28,1.44]亲[1.44,1.60]因[1.60,1.76]为[1.76,1.84]发[1.84,2.08]生[2.08,2.16]交[2.16,2.40]通[2.40,2.48]事[2.48,2.72]故[2.72,2.80]撒[2.88,3.04]手[3.04,3.20]人[3.20,3.36]寰[3.36,3.52]。桂[3.52,3.68]兰[3.68,3.84]的[3.84,4.00]母[4.00,4.08]亲[4.08,4.24]敲[4.24,4.48]了[4.48,4.56]几[4.56,4.72]下[4.72,4.80]杯[4.80,5.04]子[5.04,5.20]......[omitted]......几[22.64,22.80]个[22.80,22.96]人[22.96,23.04]聊[23.04,23.28]的[23.28,23.36]还[23.36,23.52]是[23.52,23.60]其[23.60,23.84]乐[23.84,24.00]融[24.00,24.08]融[24.08,24.24]。他[24.32,24.48]们[24.48,24.56]非[24.64,24.80]常[24.80,24.96]欢[24.96,25.12]迎[25.12,25.20]铁[25.28,25.44]柱[25.44,25.52]到[25.52,25.60]这[25.60,25.76]里[25.76,25.84]做[25.84,26.00]客[26.00,26.16]，但[26.32,26.48]这[26.48,26.64]个[26.64,26.72]女[26.72,26.96]佣[26.96,27.04]却[27.04,27.20]显[27.20,27.36]得[27.36,27.52]异[27.52,27.68]常[27.68,27.84]诡[27.84,28.00]异[28.00,28.16]。