README.md · Qwen3-ASR-1.7B

1

---

2

license: apache-2.0

3

pipeline_tag: automatic-speech-recognition

4

---

5

6

# Qwen3-ASR

7

8

## Overview

9

10

### Introduction

11

12

13

<img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/qwen3_asr_introduction.png" width="90%"/>

14

15

16

The Qwen3-ASR family includes Qwen3-ASR-1.7B and Qwen3-ASR-0.6B, which support language identification and ASR for 52 languages and dialects. Both leverage large-scale speech training data and the strong audio understanding capability of their foundation model, Qwen3-Omni. Experiments show that the 1.7B version achieves state-of-the-art performance among open-source ASR models and is competitive with the strongest proprietary commercial APIs. Here are the main features:

17

18

* **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.

19

20

* **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 long audio.

21

22

* **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.

23

24

* **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.

25

26

### Model Architecture

27

28

29

<img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/overview.jpg" width="100%"/>

30

31

32

33

### Released Models Description and Download

34

35

Below is an introduction and download information for the Qwen3-ASR models. Please select and download the model that fits your needs.

36

37

38

|---|---|---|---|---|

39

| 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 |

40

| Qwen3-ForcedAligner-0.6B | Chinese, English, Cantonese, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish | -- | NAR | Speech |

41

42

During model loading in the `qwen-asr` package or vLLM, model weights will be downloaded automatically based on the model name. However, if your runtime environment does not allow downloading weights during execution, you can use the following commands to manually download the model weights to a local directory:

43

44

```bash

45

# Download through ModelScope (recommended for users in Mainland China)

46

pip install -U modelscope

47

modelscope download --model Qwen/Qwen3-ASR-1.7B --local_dir ./Qwen3-ASR-1.7B

48

modelscope download --model Qwen/Qwen3-ASR-0.6B --local_dir ./Qwen3-ASR-0.6B

49

modelscope download --model Qwen/Qwen3-ForcedAligner-0.6B --local_dir ./Qwen3-ForcedAligner-0.6B

50

# Download through Hugging Face

51

pip install -U "huggingface_hub[cli]"

52

huggingface-cli download Qwen/Qwen3-ASR-1.7B --local-dir ./Qwen3-ASR-1.7B

53

huggingface-cli download Qwen/Qwen3-ASR-0.6B --local-dir ./Qwen3-ASR-0.6B

54

huggingface-cli download Qwen/Qwen3-ForcedAligner-0.6B --local-dir ./Qwen3-ForcedAligner-0.6B

55

```

56

57

58

## Quickstart

59

60

### Environment Setup

61

62

The easiest way to use Qwen3-ASR is to install the `qwen-asr` Python package from PyPI. This will pull in the required runtime dependencies and allow you to load any released Qwen3-ASR model. If you’d like to simplify environment setup further, you can also use our official [Docker image](#docker). The `qwen-asr` package provides two backends: the transformers backend and the vLLM backend. For usage instructions for different backends, please refer to [Python Package Usage](#python-package-usage). We recommend using a **fresh, isolated environment** to avoid dependency conflicts with existing packages. You can create a clean Python 3.12 environment like this:

63

64

```bash

65

conda create -n qwen3-asr python=3.12 -y

66

conda activate qwen3-asr

67

```

68

69

Run the following command to get the minimal installation with transformers-backend support:

70

71

```bash

72

pip install -U qwen-asr

73

```

74

75

To enable the vLLM backend for faster inference and streaming support, run:

76

77

```bash

78

pip install -U qwen-asr[vllm]

79

```

80

81

If you want to develop or modify the code locally, install from source in editable mode:

82

83

```bash

84

git clone https://github.com/QwenLM/Qwen3-ASR.git

85

cd Qwen3-ASR

86

pip install -e .

87

# support vLLM backend

88

# pip install -e ".[vllm]"

89

```

90

91

Additionally, we recommend using FlashAttention 2 to reduce GPU memory usage and accelerate inference speed, especially for long inputs and large batch sizes.

92

93

```bash

94

pip install -U flash-attn --no-build-isolation

95

```

96

97

If your machine has less than 96GB of RAM and lots of CPU cores, run:

98

99

```bash

100

MAX_JOBS=4 pip install -U flash-attn --no-build-isolation

101

```

102

103

Also, you should have hardware that is compatible with FlashAttention 2. Read more about it in the official documentation of the [FlashAttention repository](https://github.com/Dao-AILab/flash-attention). FlashAttention 2 can only be used when a model is loaded in `torch.float16` or `torch.bfloat16`.

104

105

### Python Package Usage

106

107

#### Quick Inference

108

109

The `qwen-asr` package provides two backends: **transformers backend** and **vLLM backend**. You can pass audio inputs as a local path, a URL, base64 data, or a `(np.ndarray, sr)` tuple, and run batch inference. To quickly try Qwen3-ASR, you can use `Qwen3ASRModel.from_pretrained(...)` for the transformers backend with the following code:

110

111

```python

112

import torch

113

from qwen_asr import Qwen3ASRModel

114

115

model = Qwen3ASRModel.from_pretrained(

116

"Qwen/Qwen3-ASR-1.7B",

117

dtype=torch.bfloat16,

118

device_map="cuda:0",

119

# attn_implementation="flash_attention_2",

120

    max_inference_batch_size=32, # Batch size limit for inference. -1 means unlimited. Smaller values can help avoid OOM.

121

max_new_tokens=256, # Maximum number of tokens to generate. Set a larger value for long audio input.

122

)

123

124

results = model.transcribe(

125

audio="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav",

126

language=None, # set "English" to force the language

127

)

128

129

print(results[0].language)

130

print(results[0].text)

131

```

132

133

If you want to return timestamps, pass `forced_aligner` and its init kwargs. Here is an example of batch inference with timestamps output:

134

135

```python

136

import torch

137

from qwen_asr import Qwen3ASRModel

138

139

model = Qwen3ASRModel.from_pretrained(

140

"Qwen/Qwen3-ASR-1.7B",

141

dtype=torch.bfloat16,

142

device_map="cuda:0",

143

# attn_implementation="flash_attention_2",

144

    max_inference_batch_size=32, # Batch size limit for inference. -1 means unlimited. Smaller values can help avoid OOM.

145

max_new_tokens=256, # Maximum number of tokens to generate. Set a larger value for long audio input.

146

forced_aligner="Qwen/Qwen3-ForcedAligner-0.6B",

147

forced_aligner_kwargs=dict(

148

dtype=torch.bfloat16,

149

device_map="cuda:0",

150

# attn_implementation="flash_attention_2",

151

),

152

)

153

154

results = model.transcribe(

155

audio=[

156

"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_zh.wav",

157

"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav",

158

],

159

language=["Chinese", "English"], # can also be set to None for automatic language detection

160

return_time_stamps=True,

161

)

162

163

for r in results:

164

print(r.language, r.text, r.time_stamps[0])

165

```

166

167

For more detailed usage examples, please refer to the [example code](https://github.com/QwenLM/Qwen3-ASR/blob/main/examples/example_qwen3_asr_transformers.py) for the transformers backend.

168

169

#### vLLM Backend

170

171

If you want the fastest inference speed with Qwen3-ASR, we strongly recommend using the vLLM backend by initializing the model with `Qwen3ASRModel.LLM(...)`. Example code is provided below. Note that you must install it via `pip install -U qwen-asr[vllm]`. If you want the model to output timestamps, it’s best to install FlashAttention via `pip install -U flash-attn --no-build-isolation` to speed up inference for the forced aligner model. Remember to wrap your code under `if __name__ == '__main__':` to avoid the `spawn` error described in [vLLM Troubleshooting](https://docs.vllm.ai/en/latest/usage/troubleshooting/#python-multiprocessing).

172

173

```python

174

import torch

175

from qwen_asr import Qwen3ASRModel

176

177

if __name__ == '__main__':

178

model = Qwen3ASRModel.LLM(

179

model="Qwen/Qwen3-ASR-1.7B",

180

gpu_memory_utilization=0.7,

181

        max_inference_batch_size=128, # Batch size limit for inference. -1 means unlimited. Smaller values can help avoid OOM.

182

max_new_tokens=4096, # Maximum number of tokens to generate. Set a larger value for long audio input.

183

forced_aligner="Qwen/Qwen3-ForcedAligner-0.6B",

184

forced_aligner_kwargs=dict(

185

dtype=torch.bfloat16,

186

device_map="cuda:0",

187

# attn_implementation="flash_attention_2",

188

),

189

)

190

191

results = model.transcribe(

192

audio=[

193

"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_zh.wav",

194

"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav",

195

],

196

language=["Chinese", "English"], # can also be set to None for automatic language detection

197

return_time_stamps=True,

198

)

199

200

for r in results:

201

print(r.language, r.text, r.time_stamps[0])

202

```

203

204

For more detailed usage examples, please refer to the [example code](https://github.com/QwenLM/Qwen3-ASR/blob/main/examples/example_qwen3_asr_vllm.py) for the vLLM backend. In addition, you can start a vLLM server via the `qwen-asr-serve` command, which is a wrapper around `vllm serve`. You can pass any arguments supported by `vllm serve`, for example:

205

206

```bash

207

qwen-asr-serve Qwen/Qwen3-ASR-1.7B --gpu-memory-utilization 0.8 --host 0.0.0.0 --port 8000

208

```

209

210

And send requests to the server via:

211

212

```python

213

import requests

214

215

url = "http://localhost:8000/v1/chat/completions"

216

headers = {"Content-Type": "application/json"}

217

218

data = {

219

"messages": [

220

{

221

"role": "user",

222

"content": [

223

{

224

"type": "audio_url",

225

"audio_url": {

226

"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav"

227

},

228

}

229

],

230

}

231

]

232

}

233

234

response = requests.post(url, headers=headers, json=data, timeout=300)

235

response.raise_for_status()

236

content = response.json()['choices'][0]['message']['content']

237

print(content)

238

239

# parse ASR output if you want

240

from qwen_asr import parse_asr_output

241

language, text = parse_asr_output(content)

242

print(language)

243

print(text)

244

```

245

246

#### Streaming Inference

247

248

Qwen3-ASR fully supports streaming inference. Currently, streaming inference is only available with the vLLM backend. Note that streaming inference does not support batch inference or returning timestamps. Please refer to the [example code](https://github.com/QwenLM/Qwen3-ASR/blob/main/examples/example_qwen3_asr_vllm_streaming.py) for details. You can also launch a streaming web demo through the [guide](#streaming-demo) to experience Qwen3-ASR’s streaming transcription capabilities.

249

250

#### ForcedAligner Usage

251

252

`Qwen3-ForcedAligner-0.6B` can align text–speech pairs and return word or character level timestamps. Here is an example of using the forced aligner directly:

253

254

```python

255

import torch

256

from qwen_asr import Qwen3ForcedAligner

257

258

model = Qwen3ForcedAligner.from_pretrained(

259

"Qwen/Qwen3-ForcedAligner-0.6B",

260

dtype=torch.bfloat16,

261

device_map="cuda:0",

262

# attn_implementation="flash_attention_2",

263

)

264

265

results = model.align(

266

audio="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_zh.wav",

267

text="甚至出现交易几乎停滞的情况。",

268

language="Chinese",

269

)

270

271

print(results[0])

272

print(results[0][0].text, results[0][0].start_time, results[0][0].end_time)

273

```

274

275

In addition, the forced aligner supports local paths / URLs / base64 data / `(np.ndarray, sr)` inputs and batch inference. Please refer to the [example code](https://github.com/QwenLM/Qwen3-ASR/blob/main/examples/example_qwen3_forced_aligner.py) for details.

276

277

### DashScope API Usage

278

279

To further explore Qwen3-ASR, we encourage you to try our DashScope API for a faster and more efficient experience. For detailed API information and documentation, please refer to the following:

280

281

| API Description | API Documentation (Mainland China) | API Documentation (International) |

282

|------------------|-----------------------------------|------------------------------------|

283

| Real-time API for Qwen3-ASR. | [https://help.aliyun.com/zh/model-studio/qwen-real-time-speech-recognition](https://help.aliyun.com/zh/model-studio/qwen-real-time-speech-recognition) | [https://www.alibabacloud.com/help/en/model-studio/qwen-real-time-speech-recognition](https://www.alibabacloud.com/help/en/model-studio/qwen-real-time-speech-recognition) |

284

| FileTrans API for Qwen3-ASR. | [https://help.aliyun.com/zh/model-studio/qwen-speech-recognition](https://help.aliyun.com/zh/model-studio/qwen-speech-recognition) | [https://www.alibabacloud.com/help/en/model-studio/qwen-speech-recognition](https://www.alibabacloud.com/help/en/model-studio/qwen-speech-recognition) |

285

286

287

## Launch Local Web UI Demo

288

289

### Gradio Demo

290

291

To launch the Qwen3-ASR web UI gradio demo, install the `qwen-asr` package and run `qwen-asr-demo`. Use the command below for help:

292

293

```bash

294

qwen-asr-demo --help

295

```

296

297

To launch the demo, you can use the following commands:

298

299

```bash

300

# Transformers backend

301

qwen-asr-demo \

302

--asr-checkpoint Qwen/Qwen3-ASR-1.7B \

303

--backend transformers \

304

--cuda-visible-devices 0 \

305

--ip 0.0.0.0 --port 8000

306

307

# Transformers backend + Forced Aligner (enable timestamps)

308

qwen-asr-demo \

309

--asr-checkpoint Qwen/Qwen3-ASR-1.7B \

310

--aligner-checkpoint Qwen/Qwen3-ForcedAligner-0.6B \

311

--backend transformers \

312

--cuda-visible-devices 0 \

313

--backend-kwargs '{"device_map":"cuda:0","dtype":"bfloat16","max_inference_batch_size":8,"max_new_tokens":256}' \

314

--aligner-kwargs '{"device_map":"cuda:0","dtype":"bfloat16"}' \

315

--ip 0.0.0.0 --port 8000

316

317

# vLLM backend + Forced Aligner (enable timestamps)

318

qwen-asr-demo \

319

--asr-checkpoint Qwen/Qwen3-ASR-1.7B \

320

--aligner-checkpoint Qwen/Qwen3-ForcedAligner-0.6B \

321

--backend vllm \

322

--cuda-visible-devices 0 \

323

--backend-kwargs '{"gpu_memory_utilization":0.7,"max_inference_batch_size":8,"max_new_tokens":2048}' \

324

--aligner-kwargs '{"device_map":"cuda:0","dtype":"bfloat16"}' \

325

--ip 0.0.0.0 --port 8000

326

```

327

328

Then open `http://<your-ip>:8000`, or access it via port forwarding in tools like VS Code.

329

330

#### Backend Notes

331

332

This demo supports two backends: transformers and vLLM. All backend-specific initialization parameters should be passed via `--backend-kwargs` as a JSON dict. If not provided, the demo will use sensible defaults.

333

334

```bash

335

# Example: override transformers init args without flash attention

336

--backend-kwargs '{"device_map":"cuda:0","dtype":"bfloat16"}'

337

338

# Example: override vLLM init args with 65% GPU memory

339

--backend-kwargs '{"gpu_memory_utilization":0.65}'

340

```

341

342

#### CUDA Device Notes

343

344

Because vLLM does not follow `cuda:0` style device selection, this demo selects GPUs by setting `CUDA_VISIBLE_DEVICES` via `--cuda-visible-devices`.

345

346

```bash

347

# Use GPU 0

348

--cuda-visible-devices 0

349

350

# Use GPU 1

351

--cuda-visible-devices 1

352

```

353

354

#### Timestamps Notes

355

356

Timestamps are only available when `--aligner-checkpoint` is provided. If you launch the demo without a forced aligner, the timestamps UI will be hidden automatically.

357

358

```bash

359

# No forced aligner

360

qwen-asr-demo --asr-checkpoint Qwen/Qwen3-ASR-1.7B

361

362

# With forced aligner

363

qwen-asr-demo \

364

--asr-checkpoint Qwen/Qwen3-ASR-1.7B \

365

--aligner-checkpoint Qwen/Qwen3-ForcedAligner-0.6B

366

```

367

368

#### HTTPS Notes

369

370

To avoid browser microphone permission issues after deploying the server, it is recommended/required to run the gradio service over HTTPS (especially when accessed remotely or behind modern browsers/gateways). Use `--ssl-certfile` and `--ssl-keyfile` to enable HTTPS. First, generate a private key and a self-signed certificate (valid for 365 days):

371

372

```bash

373

openssl req -x509 -newkey rsa:2048 \

374

-keyout key.pem -out cert.pem \

375

-days 365 -nodes \

376

-subj "/CN=localhost"

377

```

378

379

Then run the demo with HTTPS:

380

381

```bash

382

qwen-asr-demo \

383

--asr-checkpoint Qwen/Qwen3-ASR-1.7B \

384

--backend transformers \

385

--cuda-visible-devices 0 \

386

--ip 0.0.0.0 --port 8000 \

387

--ssl-certfile cert.pem \

388

--ssl-keyfile key.pem \

389

--no-ssl-verify

390

```

391

392

Then open `https://<your-ip>:8000` to use it. If your browser shows a warning, that’s expected for self-signed certificates. For production, use a real certificate.

393

394

### Streaming Demo

395

396

To experience Qwen3-ASR’s streaming transcription capability in a web UI, we provide a minimal Flask-based streaming demo. The demo captures microphone audio in the browser, resamples it to 16,000 Hz, and continuously pushes PCM chunks to the model. Run the demo with the following command:

397

398

```bash

399

qwen-asr-demo-streaming \

400

--asr-model-path Qwen/Qwen3-ASR-1.7B \

401

--host 0.0.0.0 \

402

--port 8000 \

403

--gpu-memory-utilization 0.9

404

```

405

406

Then open `http://<your-ip>:8000`, or access it via port forwarding in tools like VS Code.

407

408

## Deployment with vLLM

409

410

vLLM officially provides day-0 model support for Qwen3-ASR for efficient inference.

411

412

### Installation

413

You can run Qwen3-ASR with vLLM nightly wheel or docker image. To install the nightly version of vLLM, we recommend using `uv` as the environment manager

414

```bash

415

uv venv

416

source .venv/bin/activate

417

uv pip install -U vllm --pre \

418

--extra-index-url https://wheels.vllm.ai/nightly/cu129 \

419

--extra-index-url https://download.pytorch.org/whl/cu129 \

420

--index-strategy unsafe-best-match

421

uv pip install "vllm[audio]" # For additional audio dependencies

422

```

423

424

### Online Serving

425

You can easily deploy Qwen3-ASR with vLLM by running the following command

426

```bash

427

vllm serve Qwen/Qwen3-ASR-1.7B

428

```

429

After the model server is successfully deployed, you can interact with it in multiple ways.

430

431

#### Using OpenAI SDK

432

```python

433

import base64

434

import httpx

435

from openai import OpenAI

436

437

# Initialize client

438

client = OpenAI(

439

base_url="http://localhost:8000/v1",

440

api_key="EMPTY"

441

)

442

443

# Create multimodal chat completion request

444

response = client.chat.completions.create(

445

model="Qwen/Qwen3-ASR-1.7B",

446

messages=[

447

{

448

"role": "user",

449

"content": [

450

{

451

"type": "audio_url",

452

"audio_url": {

453

{"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav"}

454

}

455

}

456

]

457

}

458

],

459

)

460

461

print(response.choices[0].message.content)

462

```

463

This model is also supported on vLLM with OpenAI transcription API.

464

```python

465

import httpx

466

from openai import OpenAI

467

468

# Initialize client

469

client = OpenAI(

470

base_url="http://localhost:8000/v1",

471

api_key="EMPTY"

472

)

473

audio_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav"

474

audio_file = httpx.get(audio_url).content

475

476

transcription = client.audio.transcriptions.create(

477

model="Qwen/Qwen3-ASR-1.7B",

478

file=audio_file,

479

)

480

481

print(transcription.text)

482

```

483

484

#### Using cURL

485

```bash

486

curl http://localhost:8000/v1/chat/completions \

487

-H "Content-Type: application/json" \

488

-d '{

489

"messages": [

490

{"role": "user", "content": [

491

        {"type": "audio_url", "audio_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav"}}

492

]}

493

]

494

}'

495

```

496

497

### Offline Inference

498

See the following example on using vLLM to run offline inference with Qwen3-ASR

499

```python

500

from vllm import LLM, SamplingParams

501

from vllm.assets.audio import AudioAsset

502

import base64

503

import requests

504

505

# Initialize the LLM

506

llm = LLM(

507

model="Qwen/Qwen3-ASR-1.7B"

508

)

509

510

# Load audio

511

audio_asset = AudioAsset("winning_call")

512

513

# Create conversation with audio content

514

conversation = [

515

{

516

"role": "user",

517

"content": [

518

{

519

"type": "audio_url",

520

"audio_url": {"url": audio_asset.url}

521

}

522

]

523

}

524

]

525

526

sampling_params = SamplingParams(temperature=0.01, max_tokens=256)

527

528

# Run inference using .chat()

529

outputs = llm.chat(conversation, sampling_params=sampling_params)

530

print(outputs[0].outputs[0].text)

531

```

532

533

534

## Docker

535

536

To make it easier to use our `qwen-asr` Python package, we provide a pre-built Docker image: [qwenllm/qwen3-asr](https://hub.docker.com/r/qwenllm/qwen3-asr). You only need to install the GPU driver and download the model files to run the code. Please follow the [NVIDIA Container Toolkit installation guide](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) to ensure Docker can access your GPU. If you are in Mainland China and have trouble reaching Docker Hub, you may use a registry mirror to accelerate image pulls.

537

538

First, pull the image and start a container:

539

540

```bash

541

LOCAL_WORKDIR=/path/to/your/workspace

542

HOST_PORT=8000

543

CONTAINER_PORT=80

544

docker run --gpus all --name qwen3-asr \

545

-v /var/run/docker.sock:/var/run/docker.sock -p $HOST_PORT:$CONTAINER_PORT \

546

--mount type=bind,source=$LOCAL_WORKDIR,target=/data/shared/Qwen3-ASR \

547

--shm-size=4gb \

548

-it qwenllm/qwen3-asr:latest

549

```

550

551

After running the command, you will enter the container’s bash shell. Your local workspace (**replace** `/path/to/your/workspace` **with the actual path**) will be mounted inside the container at `/data/shared/Qwen3-ASR`. Port `8000` on the host is mapped to port `80` in the container, so you can access services running in the container via `http://<host-ip>:8000`. Note that services inside the container must bind to `0.0.0.0` (not `127.0.0.1`) for port forwarding to work.

552

553

If you exit the container, you can start it again and re-enter it with:

554

555

```bash

556

docker start qwen3-asr

557

docker exec -it qwen3-asr bash

558

```

559

560

To remove the container completely, run:

561

562

```bash

563

docker rm -f qwen3-asr

564

```

565

566

567

## Evaluation

568

569

During evaluation, we ran inference for all models with `dtype=torch.bfloat16` and set `max_new_tokens=1024` using vLLM. Greedy search was used for all decoding, and none of the tests specified a language parameter. The detailed evaluation results are shown below.

570

571

<details>

572

<summary>ASR Benchmarks on Public Datasets (WER ↓)</summary>

573

574

<table>

575

<thead>

576

<tr>

577

<th colspan="2" style="text-align: left;"></th>

578

<th style="text-align: center;">GPT-4o -Transcribe</th>

579

<th style="text-align: center;">Gemini-2.5 -Pro</th>

580

<th style="text-align: center;">Doubao-ASR</th>

581

<th style="text-align: center;">Whisper -large-v3</th>

582

<th style="text-align: center;">Fun-ASR -MLT-Nano</th>

583

<th style="text-align: center;">Qwen3-ASR -0.6B</th>

584

<th style="text-align: center;">Qwen3-ASR -1.7B</th>

585

</tr>

586

</thead>

587

<tbody>

588

<tr>

589

      <td colspan="9" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">English (en)</td>

590

</tr>

591

<tr>

592

<td colspan="2" style="text-align: left;">Librispeech clean | other</td>

593

<td style="text-align: center;">1.39 | 3.75</td>

594

<td style="text-align: center;">2.89 | 3.56</td>

595

<td style="text-align: center;">2.78 | 5.70</td>

596

<td style="text-align: center;">1.51 | 3.97</td>

597

<td style="text-align: center;">1.68 | 4.03</td>

598

<td style="text-align: center;">2.11 | 4.55</td>

599

<td style="text-align: center;">1.63 | 3.38</td>

600

</tr>

601

<tr>

602

<td colspan="2" style="text-align: left;">GigaSpeech</td>

603

<td style="text-align: center;">25.50</td>

604

<td style="text-align: center;">9.37</td>

605

<td style="text-align: center;">9.55</td>

606

<td style="text-align: center;">9.76</td>

607

<td style="text-align: center;">-</td>

608

<td style="text-align: center;">8.88</td>

609

<td style="text-align: center;">8.45</td>

610

</tr>

611

<tr>

612

<td colspan="2" style="text-align: left;">CV-en</td>

613

<td style="text-align: center;">9.08</td>

614

<td style="text-align: center;">14.49</td>

615

<td style="text-align: center;">13.78</td>

616

<td style="text-align: center;">9.90</td>

617

<td style="text-align: center;">9.90</td>

618

<td style="text-align: center;">9.92</td>

619

<td style="text-align: center;">7.39</td>

620

</tr>

621

<tr>

622

<td colspan="2" style="text-align: left;">Fleurs-en</td>

623

<td style="text-align: center;">2.40</td>

624

<td style="text-align: center;">2.94</td>

625

<td style="text-align: center;">6.31</td>

626

<td style="text-align: center;">4.08</td>

627

<td style="text-align: center;">5.49</td>

628

<td style="text-align: center;">4.39</td>

629

<td style="text-align: center;">3.35</td>

630

</tr>

631

<tr>

632

<td colspan="2" style="text-align: left;">MLS-en</td>

633

<td style="text-align: center;">5.12</td>

634

<td style="text-align: center;">3.68</td>

635

<td style="text-align: center;">7.09</td>

636

<td style="text-align: center;">4.87</td>

637

<td style="text-align: center;">-</td>

638

<td style="text-align: center;">6.00</td>

639

<td style="text-align: center;">4.58</td>

640

</tr>

641

<tr>

642

<td colspan="2" style="text-align: left;">Tedlium</td>

643

<td style="text-align: center;">7.69</td>

644

<td style="text-align: center;">6.15</td>

645

<td style="text-align: center;">4.91</td>

646

<td style="text-align: center;">6.84</td>

647

<td style="text-align: center;">-</td>

648

<td style="text-align: center;">3.85</td>

649

<td style="text-align: center;">4.50</td>

650

</tr>

651

<tr>

652

<td colspan="2" style="text-align: left;">VoxPopuli</td>

653

<td style="text-align: center;">10.29</td>

654

<td style="text-align: center;">11.36</td>

655

<td style="text-align: center;">12.12</td>

656

<td style="text-align: center;">12.05</td>

657

<td style="text-align: center;">-</td>

658

<td style="text-align: center;">9.96</td>

659

<td style="text-align: center;">9.15</td>

660

</tr>

661

<tr>

662

      <td colspan="9" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Chinese (zh)</td>

663

</tr>

664

<tr>

665

<td colspan="2" style="text-align: left;">WenetSpeech net | meeting</td>

666

<td style="text-align: center;">15.30 | 32.27</td>

667

<td style="text-align: center;">14.43 | 13.47</td>

668

<td style="text-align: center;">N/A</td>

669

<td style="text-align: center;">9.86 | 19.11</td>

670

<td style="text-align: center;">6.35 | -</td>

671

<td style="text-align: center;">5.97 | 6.88</td>

672

<td style="text-align: center;">4.97 | 5.88</td>

673

</tr>

674

<tr>

675

<td colspan="2" style="text-align: left;">AISHELL-2-test</td>

676

<td style="text-align: center;">4.24</td>

677

<td style="text-align: center;">11.62</td>

678

<td style="text-align: center;">2.85</td>

679

<td style="text-align: center;">5.06</td>

680

<td style="text-align: center;">-</td>

681

<td style="text-align: center;">3.15</td>

682

<td style="text-align: center;">2.71</td>

683

</tr>

684

<tr>

685

<td colspan="2" style="text-align: left;">SpeechIO</td>

686

<td style="text-align: center;">12.86</td>

687

<td style="text-align: center;">5.30</td>

688

<td style="text-align: center;">2.93</td>

689

<td style="text-align: center;">7.56</td>

690

<td style="text-align: center;">-</td>

691

<td style="text-align: center;">3.44</td>

692

<td style="text-align: center;">2.88</td>

693

</tr>

694

<tr>

695

<td colspan="2" style="text-align: left;">Fleurs-zh</td>

696

<td style="text-align: center;">2.44</td>

697

<td style="text-align: center;">2.71</td>

698

<td style="text-align: center;">2.69</td>

699

<td style="text-align: center;">4.09</td>

700

<td style="text-align: center;">3.51</td>

701

<td style="text-align: center;">2.88</td>

702

<td style="text-align: center;">2.41</td>

703

</tr>

704

<tr>

705

<td colspan="2" style="text-align: left;">CV-zh</td>

706

<td style="text-align: center;">6.32</td>

707

<td style="text-align: center;">7.70</td>

708

<td style="text-align: center;">5.95</td>

709

<td style="text-align: center;">12.91</td>

710

<td style="text-align: center;">6.20</td>

711

<td style="text-align: center;">6.89</td>

712

<td style="text-align: center;">5.35</td>

713

</tr>

714

<tr>

715

      <td colspan="9" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Chinese Dialect</td>

716

</tr>

717

<tr>

718

<td colspan="2" style="text-align: left;">KeSpeech</td>

719

<td style="text-align: center;">26.87</td>

720

<td style="text-align: center;">24.71</td>

721

<td style="text-align: center;">5.27</td>

722

<td style="text-align: center;">28.79</td>

723

<td style="text-align: center;">-</td>

724

<td style="text-align: center;">7.08</td>

725

<td style="text-align: center;">5.10</td>

726

</tr>

727

<tr>

728

<td colspan="2" style="text-align: left;">Fleurs-yue</td>

729

<td style="text-align: center;">4.98</td>

730

<td style="text-align: center;">9.43</td>

731

<td style="text-align: center;">4.98</td>

732

<td style="text-align: center;">9.18</td>

733

<td style="text-align: center;">-</td>

734

<td style="text-align: center;">5.79</td>

735

<td style="text-align: center;">3.98</td>

736

</tr>

737

<tr>

738

<td colspan="2" style="text-align: left;">CV-yue</td>

739

<td style="text-align: center;">11.36</td>

740

<td style="text-align: center;">18.76</td>

741

<td style="text-align: center;">13.20</td>

742

<td style="text-align: center;">16.23</td>

743

<td style="text-align: center;">-</td>

744

<td style="text-align: center;">9.50</td>

745

<td style="text-align: center;">7.57</td>

746

</tr>

747

<tr>

748

<td colspan="2" style="text-align: left;">CV-zh-tw</td>

749

<td style="text-align: center;">6.32</td>

750

<td style="text-align: center;">7.31</td>

751

<td style="text-align: center;">4.06</td>

752

<td style="text-align: center;">7.84</td>

753

<td style="text-align: center;">-</td>

754

<td style="text-align: center;">5.59</td>

755

<td style="text-align: center;">3.77</td>

756

</tr>

757

<tr>

758

<td colspan="2" style="text-align: left;">WenetSpeech-Yue short | long</td>

759

<td style="text-align: center;">15.62 | 25.29</td>

760

<td style="text-align: center;">25.19 | 11.23</td>

761

<td style="text-align: center;">9.74 | 11.40</td>

762

<td style="text-align: center;">32.26 | 46.64</td>

763

<td style="text-align: center;">- | -</td>

764

<td style="text-align: center;">7.54 | 9.92</td>

765

<td style="text-align: center;">5.82 | 8.85</td>

766

</tr>

767

<tr>

768

<td colspan="2" style="text-align: left;">WenetSpeech-Chuan easy | hard</td>

769

<td style="text-align: center;">34.81 | 53.98</td>

770

<td style="text-align: center;">43.79 | 67.30</td>

771

<td style="text-align: center;">11.40 | 20.20</td>

772

<td style="text-align: center;">14.35 | 26.80</td>

773

<td style="text-align: center;">- | -</td>

774

<td style="text-align: center;">13.92 | 24.45</td>

775

<td style="text-align: center;">11.99 | 21.63</td>

776

</tr>

777

</tbody>

778

</table>

779

780

</details>

781

782

<details>

783

<summary>ASR Benchmarks on Internal Datasets (WER ↓)</summary>

784

785

<table>

786

<thead>

787

<tr>

788

<th style="text-align: left;"></th>

789

<th style="text-align: center;">GPT-4o -Transcribe</th>

790

<th style="text-align: center;">Gemini-2.5 -Pro</th>

791

<th style="text-align: center;">Doubao-ASR</th>

792

<th style="text-align: center;">Whisper -large-v3</th>

793

<th style="text-align: center;">Fun-ASR -MLT-Nano</th>

794

<th style="text-align: center;">Qwen3-ASR -0.6B</th>

795

<th style="text-align: center;">Qwen3-ASR -1.7B</th>

796

</tr>

797

</thead>

798

<tbody>

799

<tr>

800

      <td colspan="8" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Accented English</td>

801

</tr>

802

<tr>

803

<td style="text-align: left;">Dialog-Accented English</td>

804

<td style="text-align: center;">28.56</td>

805

<td style="text-align: center;">23.85</td>

806

<td style="text-align: center;">20.41</td>

807

<td style="text-align: center;">21.30</td>

808

<td style="text-align: center;">19.96</td>

809

<td style="text-align: center;">16.62</td>

810

<td style="text-align: center;">16.07</td>

811

</tr>

812

<tr>

813

      <td colspan="8" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Chinese Mandarin</td>

814

</tr>

815

<tr>

816

<td style="text-align: left;">Elders&Kids</td>

817

<td style="text-align: center;">14.27</td>

818

<td style="text-align: center;">36.93</td>

819

<td style="text-align: center;">4.17</td>

820

<td style="text-align: center;">10.61</td>

821

<td style="text-align: center;">4.54</td>

822

<td style="text-align: center;">4.48</td>

823

<td style="text-align: center;">3.81</td>

824

</tr>

825

<tr>

826

<td style="text-align: left;">ExtremeNoise</td>

827

<td style="text-align: center;">36.11</td>

828

<td style="text-align: center;">29.06</td>

829

<td style="text-align: center;">17.04</td>

830

<td style="text-align: center;">63.17</td>

831

<td style="text-align: center;">36.55</td>

832

<td style="text-align: center;">17.88</td>

833

<td style="text-align: center;">16.17</td>

834

</tr>

835

<tr>

836

<td style="text-align: left;">TongueTwister</td>

837

<td style="text-align: center;">20.87</td>

838

<td style="text-align: center;">4.97</td>

839

<td style="text-align: center;">3.47</td>

840

<td style="text-align: center;">16.63</td>

841

<td style="text-align: center;">9.02</td>

842

<td style="text-align: center;">4.06</td>

843

<td style="text-align: center;">2.44</td>

844

</tr>

845

<tr>

846

<td style="text-align: left;">Dialog-Mandarin</td>

847

<td style="text-align: center;">20.73</td>

848

<td style="text-align: center;">12.50</td>

849

<td style="text-align: center;">6.61</td>

850

<td style="text-align: center;">14.01</td>

851

<td style="text-align: center;">7.32</td>

852

<td style="text-align: center;">7.06</td>

853

<td style="text-align: center;">6.54</td>

854

</tr>

855

<tr>

856

      <td colspan="8" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Chinese Dialect</td>

857

</tr>

858

<tr>

859

<td style="text-align: left;">Dialog-Cantonese</td>

860

<td style="text-align: center;">16.05</td>

861

<td style="text-align: center;">14.98</td>

862

<td style="text-align: center;">7.56</td>

863

<td style="text-align: center;">31.04</td>

864

<td style="text-align: center;">5.85</td>

865

<td style="text-align: center;">4.80</td>

866

<td style="text-align: center;">4.12</td>

867

</tr>

868

<tr>

869

<td style="text-align: left;">Dialog-Chinese Dialects</td>

870

<td style="text-align: center;">45.37</td>

871

<td style="text-align: center;">47.70</td>

872

<td style="text-align: center;">19.85</td>

873

<td style="text-align: center;">44.55</td>

874

<td style="text-align: center;">19.41</td>

875

<td style="text-align: center;">18.24</td>

876

<td style="text-align: center;">15.94</td>

877

</tr>

878

</tbody>

879

</table>

880

<p><strong>Dialect coverage:</strong> Results for <em>Dialog-Accented English</em> are averaged over 16 accents, and results for <em>Dialog-Chinese Dialects</em> are averaged over 22 Chinese dialects.</p>

881

882

</details>

883

884

<details>

885

<summary>Multilingual ASR Benchmarks (WER ↓)</summary>

886

887

<table>

888

<thead>

889

<tr>

890

<th style="text-align: left;"></th>

891

<th style="text-align: center;">GLM-ASR -Nano-2512</th>

892

<th style="text-align: center;">Whisper -large-v3</th>

893

<th style="text-align: center;">Fun-ASR -MLT-Nano</th>

894

<th style="text-align: center;">Qwen3-ASR -0.6B</th>

895

<th style="text-align: center;">Qwen3-ASR -1.7B</th>

896

</tr>

897

</thead>

898

<tbody>

899

<tr>

900

      <td colspan="6" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Open-sourced Benchmarks</td>

901

</tr>

902

<tr>

903

<td style="text-align: left;">MLS</td>

904

<td style="text-align: center;">13.32</td>

905

<td style="text-align: center;">8.62</td>

906

<td style="text-align: center;">28.70</td>

907

<td style="text-align: center;">13.19</td>

908

<td style="text-align: center;">8.55</td>

909

</tr>

910

<tr>

911

<td style="text-align: left;">CommonVoice</td>

912

<td style="text-align: center;">19.40</td>

913

<td style="text-align: center;">10.77</td>

914

<td style="text-align: center;">17.25</td>

915

<td style="text-align: center;">12.75</td>

916

<td style="text-align: center;">9.18</td>

917

</tr>

918

<tr>

919

<td style="text-align: left;">MLC-SLM</td>

920

<td style="text-align: center;">34.93</td>

921

<td style="text-align: center;">15.68</td>

922

<td style="text-align: center;">29.94</td>

923

<td style="text-align: center;">15.84</td>

924

<td style="text-align: center;">12.74</td>

925

</tr>

926

<tr>

927

<td style="text-align: left;">Fleurs</td>

928

<td style="text-align: center;">16.08</td>

929

<td style="text-align: center;">5.27</td>

930

<td style="text-align: center;">10.03</td>

931

<td style="text-align: center;">7.57</td>

932

<td style="text-align: center;">4.90</td>

933

</tr>

934

<tr>

935

<td style="text-align: left;">Fleurs†</td>

936

<td style="text-align: center;">20.05</td>

937

<td style="text-align: center;">6.85</td>

938

<td style="text-align: center;">31.89</td>

939

<td style="text-align: center;">10.37</td>

940

<td style="text-align: center;">6.62</td>

941

</tr>

942

<tr>

943

<td style="text-align: left;">Fleurs††</td>

944

<td style="text-align: center;">24.83</td>

945

<td style="text-align: center;">8.16</td>

946

<td style="text-align: center;">47.84</td>

947

<td style="text-align: center;">21.80</td>

948

<td style="text-align: center;">12.60</td>

949

</tr>

950

<tr>

951

      <td colspan="6" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Qwen-ASR Internal Benchmarks</td>

952

</tr>

953

<tr>

954

<td style="text-align: left;">News-Multilingual</td>

955

<td style="text-align: center;">49.40</td>

956

<td style="text-align: center;">14.80</td>

957

<td style="text-align: center;">65.07</td>

958

<td style="text-align: center;">17.39</td>

959

<td style="text-align: center;">12.80</td>

960

</tr>

961

</tbody>

962

</table>

963

<p><strong>Language coverage:</strong> <em>MLS</em> includes 8 languages: {da, de, en, es, fr, it, pl, pt}.<br><em>CommonVoice</em> includes 13 languages: {en, zh, yue, zh_TW, ar, de, es, fr, it, ja, ko, pt, ru}.<br><em>MLC-SLM</em> includes 11 languages: {en, fr, de, it, pt, es, ja, ko, ru, th, vi}.<br><em>Fleurs</em> includes 12 languages: {en, zh, yue, ar, de, es, fr, it, ja, ko, pt, ru }.<br><em>Fleurs<sup>†</sup></em> includes 8 additional languages beyond Fleurs: {hi, id, ms, nl, pl, th, tr, vi}.<br><em>Fleurs<sup>††</sup></em> includes 10 additional languages beyond Fleurs<sup>†</sup>: {cs, da, el, fa, fi, fil, hu, mk, ro, sv}.<br><em>News-Multilingual</em> includes 15 languages: {ar, de, es, fr, hi, id, it, ja, ko, nl, pl, pt, ru, th, vi}.</p>

964

965

</details>

966

967

<details>

968

<summary>Language Identification Accuracy (%) ↑</summary>

969

970

<table>

971

<thead>

972

<tr>

973

<th style="text-align: left;"></th>

974

<th style="text-align: center;">Whisper-large-v3</th>

975

<th style="text-align: center;">Qwen3-ASR-0.6B</th>

976

<th style="text-align: center;">Qwen3-ASR-1.7B</th>

977

</tr>

978

</thead>

979

<tbody>

980

<tr>

981

<td style="text-align: left;">MLS</td>

982

<td style="text-align: center;">99.9</td>

983

<td style="text-align: center;">99.3</td>

984

<td style="text-align: center;">99.9</td>

985

</tr>

986

<tr>

987

<td style="text-align: left;">CommonVoice</td>

988

<td style="text-align: center;">92.7</td>

989

<td style="text-align: center;">98.2</td>

990

<td style="text-align: center;">98.7</td>

991

</tr>

992

<tr>

993

<td style="text-align: left;">MLC-SLM</td>

994

<td style="text-align: center;">89.2</td>

995

<td style="text-align: center;">92.7</td>

996

<td style="text-align: center;">94.1</td>

997

</tr>

998

<tr>

999

<td style="text-align: left;">Fleurs</td>

1000

<td style="text-align: center;">94.6</td>

1001

<td style="text-align: center;">97.1</td>

1002

<td style="text-align: center;">98.7</td>

1003

</tr>

1004

<tr style="border-top: 1px solid #ddd;">

1005

<td style="text-align: left;">Avg.</td>

1006

<td style="text-align: center;">94.1</td>

1007

<td style="text-align: center;">96.8</td>

1008

<td style="text-align: center;">97.9</td>

1009

</tr>

1010

</tbody>

1011

</table>

1012

<p><strong>Language coverage:</strong> The language sets follow Multilingual ASR Benchmarks. Here, Fleurs corresponds to Fleurs<sup>††</sup> in Multilingual ASR Benchmarks and covers 30 languages.</p>

1013

1014

</details>

1015

1016

<details>

1017

<summary>Singing Voice & Song Transcription (WER ↓)</summary>

1018

1019

<table>

1020

<thead>

1021

<tr>

1022

<th style="text-align: left;"></th>

1023

<th style="text-align: center;">GPT-4o -Transcribe</th>

1024

<th style="text-align: center;">Gemini-2.5 -Pro</th>

1025

<th style="text-align: center;">Doubao-ASR -1.0</th>

1026

<th style="text-align: center;">Whisper -large-v3</th>

1027

<th style="text-align: center;">Fun-ASR-MLT -Nano</th>

1028

<th style="text-align: center;">Qwen3-ASR -1.7B</th>

1029

</tr>

1030

</thead>

1031

<tbody>

1032

<tr>

1033

      <td colspan="7" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Singing</td>

1034

</tr>

1035

<tr>

1036

<td style="text-align: left;">M4Singer</td>

1037

<td style="text-align: center;">16.77</td>

1038

<td style="text-align: center;">20.88</td>

1039

<td style="text-align: center;">7.88</td>

1040

<td style="text-align: center;">13.58</td>

1041

<td style="text-align: center;">7.29</td>

1042

<td style="text-align: center;">5.98</td>

1043

</tr>

1044

<tr>

1045

<td style="text-align: left;">MIR-1k-vocal</td>

1046

<td style="text-align: center;">11.87</td>

1047

<td style="text-align: center;">9.85</td>

1048

<td style="text-align: center;">6.56</td>

1049

<td style="text-align: center;">11.71</td>

1050

<td style="text-align: center;">8.17</td>

1051

<td style="text-align: center;">6.25</td>

1052

</tr>

1053

<tr>

1054

<td style="text-align: left;">Opencpop</td>

1055

<td style="text-align: center;">7.93</td>

1056

<td style="text-align: center;">6.49</td>

1057

<td style="text-align: center;">3.80</td>

1058

<td style="text-align: center;">9.52</td>

1059

<td style="text-align: center;">2.98</td>

1060

<td style="text-align: center;">3.08</td>

1061

</tr>

1062

<tr>

1063

<td style="text-align: left;">Popcs</td>

1064

<td style="text-align: center;">32.84</td>

1065

<td style="text-align: center;">15.13</td>

1066

<td style="text-align: center;">8.97</td>

1067

<td style="text-align: center;">13.77</td>

1068

<td style="text-align: center;">9.42</td>

1069

<td style="text-align: center;">8.52</td>

1070

</tr>

1071

<tr>

1072

      <td colspan="7" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Songs with BGM</td>

1073

</tr>

1074

<tr>

1075

<td style="text-align: left;">EntireSongs-en</td>

1076

<td style="text-align: center;">30.71</td>

1077

<td style="text-align: center;">12.18</td>

1078

<td style="text-align: center;">33.51</td>

1079

<td style="text-align: center;">N/A</td>

1080

<td style="text-align: center;">N/A</td>

1081

<td style="text-align: center;">14.60</td>

1082

</tr>

1083

<tr>

1084

<td style="text-align: left;">EntireSongs-zh</td>

1085

<td style="text-align: center;">34.86</td>

1086

<td style="text-align: center;">18.68</td>

1087

<td style="text-align: center;">23.99</td>

1088

<td style="text-align: center;">N/A</td>

1089

<td style="text-align: center;">N/A</td>

1090

<td style="text-align: center;">13.91</td>

1091

</tr>

1092

</tbody>

1093

</table>

1094

1095

</details>

1096

1097

<details>

1098

<summary>ASR Inference Mode Performance (WER ↓)</summary>

1099

1100

<table>

1101

<thead>

1102

<tr>

1103

<th style="text-align: left;">Model</th>

1104

<th style="text-align: left;">Infer. Mode</th>

1105

<th style="text-align: center;">Librispeech</th>

1106

<th style="text-align: center;">Fleurs-en</th>

1107

<th style="text-align: center;">Fleurs-zh</th>

1108

<th style="text-align: center;">Avg.</th>

1109

</tr>

1110

</thead>

1111

<tbody>

1112

<tr>

1113

<td rowspan="2" style="text-align: left; vertical-align: middle;">Qwen3-ASR-1.7B</td>

1114

<td style="text-align: left;">Offline</td>

1115

<td style="text-align: center;">1.63 | 3.38</td>

1116

<td style="text-align: center;">3.35</td>

1117

<td style="text-align: center;">2.41</td>

1118

<td style="text-align: center;">2.69</td>

1119

</tr>

1120

<tr>

1121

<td style="text-align: left;">Streaming</td>

1122

<td style="text-align: center;">1.95 | 4.51</td>

1123

<td style="text-align: center;">4.02</td>

1124

<td style="text-align: center;">2.84</td>

1125

<td style="text-align: center;">3.33</td>

1126

</tr>

1127

<tr style="border-top: 1px solid #ddd;">

1128

<td rowspan="2" style="text-align: left; vertical-align: middle;">Qwen3-ASR-0.6B</td>

1129

<td style="text-align: left;">Offline</td>

1130

<td style="text-align: center;">2.11 | 4.55</td>

1131

<td style="text-align: center;">4.39</td>

1132

<td style="text-align: center;">2.88</td>

1133

<td style="text-align: center;">3.48</td>

1134

</tr>

1135

<tr>

1136

<td style="text-align: left;">Streaming</td>

1137

<td style="text-align: center;">2.54 | 6.27</td>

1138

<td style="text-align: center;">5.38</td>

1139

<td style="text-align: center;">3.40</td>

1140

<td style="text-align: center;">4.40</td>

1141

</tr>

1142

</tbody>

1143

</table>

1144

1145

</details>

1146

1147

<details>

1148

<summary>Forced Alignment Benchmarks (AAS ms ↓)</summary>

1149

1150

<table>

1151

<thead>

1152

<tr>

1153

<th style="text-align: left;"></th>

1154

<th style="text-align: center;">Monotonic-Aligner</th>

1155

<th style="text-align: center;">NFA</th>

1156

<th style="text-align: center;">WhisperX</th>

1157

<th style="text-align: center;">Qwen3-ForcedAligner-0.6B</th>

1158

</tr>

1159

</thead>

1160

<tbody>

1161

<tr>

1162

      <td colspan="5" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">MFA-Labeled Raw</td>

1163

</tr>

1164

<tr>

1165

<td style="text-align: left;">Chinese</td>

1166

<td style="text-align: center;">161.1</td>

1167

<td style="text-align: center;">109.8</td>

1168

<td style="text-align: center;">-</td>

1169

<td style="text-align: center;">33.1</td>

1170

</tr>

1171

<tr>

1172

<td style="text-align: left;">English</td>

1173

<td style="text-align: center;">-</td>

1174

<td style="text-align: center;">107.5</td>

1175

<td style="text-align: center;">92.1</td>

1176

<td style="text-align: center;">37.5</td>

1177

</tr>

1178

<tr>

1179

<td style="text-align: left;">French</td>

1180

<td style="text-align: center;">-</td>

1181

<td style="text-align: center;">100.7</td>

1182

<td style="text-align: center;">145.3</td>

1183

<td style="text-align: center;">41.7</td>

1184

</tr>

1185

<tr>

1186

<td style="text-align: left;">German</td>

1187

<td style="text-align: center;">-</td>

1188

<td style="text-align: center;">122.7</td>

1189

<td style="text-align: center;">165.1</td>

1190

<td style="text-align: center;">46.5</td>

1191

</tr>

1192

<tr>

1193

<td style="text-align: left;">Italian</td>

1194

<td style="text-align: center;">-</td>

1195

<td style="text-align: center;">142.7</td>

1196

<td style="text-align: center;">155.5</td>

1197

<td style="text-align: center;">75.5</td>

1198

</tr>

1199

<tr>

1200

<td style="text-align: left;">Japanese</td>

1201

<td style="text-align: center;">-</td>

1202

<td style="text-align: center;">-</td>

1203

<td style="text-align: center;">-</td>

1204

<td style="text-align: center;">42.2</td>

1205

</tr>

1206

<tr>

1207

<td style="text-align: left;">Korean</td>

1208

<td style="text-align: center;">-</td>

1209

<td style="text-align: center;">-</td>

1210

<td style="text-align: center;">-</td>

1211

<td style="text-align: center;">37.2</td>

1212

</tr>

1213

<tr>

1214

<td style="text-align: left;">Portuguese</td>

1215

<td style="text-align: center;">-</td>

1216

<td style="text-align: center;">-</td>

1217

<td style="text-align: center;">-</td>

1218

<td style="text-align: center;">38.4</td>

1219

</tr>

1220

<tr>

1221

<td style="text-align: left;">Russian</td>

1222

<td style="text-align: center;">-</td>

1223

<td style="text-align: center;">200.7</td>

1224

<td style="text-align: center;">-</td>

1225

<td style="text-align: center;">40.2</td>

1226

</tr>

1227

<tr>

1228

<td style="text-align: left;">Spanish</td>

1229

<td style="text-align: center;">-</td>

1230

<td style="text-align: center;">124.7</td>

1231

<td style="text-align: center;">108.0</td>

1232

<td style="text-align: center;">36.8</td>

1233

</tr>

1234

<tr>

1235

<td style="text-align: left;">Avg.</td>

1236

<td style="text-align: center;">161.1</td>

1237

<td style="text-align: center;">129.8</td>

1238

<td style="text-align: center;">133.2</td>

1239

<td style="text-align: center;">42.9</td>

1240

</tr>

1241

<tr>

1242

      <td colspan="5" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">MFA-Labeled Concat-300s</td>

1243

</tr>

1244

<tr>

1245

<td style="text-align: left;">Chinese</td>

1246

<td style="text-align: center;">1742.4</td>

1247

<td style="text-align: center;">235.0</td>

1248

<td style="text-align: center;">-</td>

1249

<td style="text-align: center;">36.5</td>

1250

</tr>

1251

<tr>

1252

<td style="text-align: left;">English</td>

1253

<td style="text-align: center;">-</td>

1254

<td style="text-align: center;">226.7</td>

1255

<td style="text-align: center;">227.2</td>

1256

<td style="text-align: center;">58.6</td>

1257

</tr>

1258

<tr>

1259

<td style="text-align: left;">French</td>

1260

<td style="text-align: center;">-</td>

1261

<td style="text-align: center;">230.6</td>

1262

<td style="text-align: center;">2052.2</td>

1263

<td style="text-align: center;">53.4</td>

1264

</tr>

1265

<tr>

1266

<td style="text-align: left;">German</td>

1267

<td style="text-align: center;">-</td>

1268

<td style="text-align: center;">220.3</td>

1269

<td style="text-align: center;">993.4</td>

1270

<td style="text-align: center;">62.4</td>

1271

</tr>

1272

<tr>

1273

<td style="text-align: left;">Italian</td>

1274

<td style="text-align: center;">-</td>

1275

<td style="text-align: center;">290.5</td>

1276

<td style="text-align: center;">5719.4</td>

1277

<td style="text-align: center;">81.6</td>

1278

</tr>

1279

<tr>

1280

<td style="text-align: left;">Japanese</td>

1281

<td style="text-align: center;">-</td>

1282

<td style="text-align: center;">-</td>

1283

<td style="text-align: center;">-</td>

1284

<td style="text-align: center;">81.3</td>

1285

</tr>

1286

<tr>

1287

<td style="text-align: left;">Korean</td>

1288

<td style="text-align: center;">-</td>

1289

<td style="text-align: center;">-</td>

1290

<td style="text-align: center;">-</td>

1291

<td style="text-align: center;">42.2</td>

1292

</tr>

1293

<tr>

1294

<td style="text-align: left;">Portuguese</td>

1295

<td style="text-align: center;">-</td>

1296

<td style="text-align: center;">-</td>

1297

<td style="text-align: center;">-</td>

1298

<td style="text-align: center;">50.0</td>

1299

</tr>

1300

<tr>

1301

<td style="text-align: left;">Russian</td>

1302

<td style="text-align: center;">-</td>

1303

<td style="text-align: center;">283.3</td>

1304

<td style="text-align: center;">-</td>

1305

<td style="text-align: center;">43.0</td>

1306

</tr>

1307

<tr>

1308

<td style="text-align: left;">Spanish</td>

1309

<td style="text-align: center;">-</td>

1310

<td style="text-align: center;">240.2</td>

1311

<td style="text-align: center;">4549.9</td>

1312

<td style="text-align: center;">39.6</td>

1313

</tr>

1314

<tr>

1315

<td style="text-align: left;">Cross-lingual</td>

1316

<td style="text-align: center;">-</td>

1317

<td style="text-align: center;">-</td>

1318

<td style="text-align: center;">-</td>

1319

<td style="text-align: center;">34.2</td>

1320

</tr>

1321

<tr>

1322

<td style="text-align: left;">Avg.</td>

1323

<td style="text-align: center;">1742.4</td>

1324

<td style="text-align: center;">246.7</td>

1325

<td style="text-align: center;">2708.4</td>

1326

<td style="text-align: center;">52.9</td>

1327

</tr>

1328

<tr>

1329

      <td colspan="5" style="text-align: left; font-style: italic; border-top: 1px solid #ddd; border-bottom: 1px solid #ddd;">Human-Labeled</td>

1330

</tr>

1331

<tr>

1332

<td style="text-align: left;">Raw</td>

1333

<td style="text-align: center;">49.9</td>

1334

<td style="text-align: center;">88.6</td>

1335

<td style="text-align: center;">-</td>

1336

<td style="text-align: center;">27.8</td>

1337

</tr>

1338

<tr>

1339

<td style="text-align: left;">Raw-Noisy</td>

1340

<td style="text-align: center;">53.3</td>

1341

<td style="text-align: center;">89.5</td>

1342

<td style="text-align: center;">-</td>

1343

<td style="text-align: center;">41.8</td>

1344

</tr>

1345

<tr>

1346

<td style="text-align: left;">Concat-60s</td>

1347

<td style="text-align: center;">51.1</td>

1348

<td style="text-align: center;">86.7</td>

1349

<td style="text-align: center;">-</td>

1350

<td style="text-align: center;">25.3</td>

1351

</tr>

1352

<tr>

1353

<td style="text-align: left;">Concat-300s</td>

1354

<td style="text-align: center;">410.8</td>

1355

<td style="text-align: center;">140.0</td>

1356

<td style="text-align: center;">-</td>

1357

<td style="text-align: center;">24.8</td>

1358

</tr>

1359

<tr>

1360

<td style="text-align: left;">Concat-Cross-lingual</td>

1361

<td style="text-align: center;">-</td>

1362

<td style="text-align: center;">-</td>

1363

<td style="text-align: center;">-</td>

1364

<td style="text-align: center;">42.5</td>

1365

</tr>

1366

<tr>

1367

<td style="text-align: left;">Avg.</td>

1368

<td style="text-align: center;">141.3</td>

1369

<td style="text-align: center;">101.2</td>

1370

<td style="text-align: center;">-</td>

1371

<td style="text-align: center;">32.4</td>

1372

</tr>

1373

</tbody>

1374

</table>

1375

1376

</details>

1377

1378

1379

## Citation

1380

1381

If you find our paper and code useful in your research, please consider giving a star :star: and citation :pencil: :)

1382

1383

```BibTeX

1384

@article{Qwen3-ASR,

1385

title={Qwen3-ASR Technical Report},

1386

  author={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},

1387

journal={arXiv preprint arXiv:2601.21337},

1388

year={2026}

1389

}

1390

```

1391

1392

1393

 