Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
feat: add Qwen3-ASR model support via AutoModel
游雁8d3e77a6b9ae
diff --git a/funasr/models/qwen3_asr/model.py b/funasr/models/qwen3_asr/model.pyindex c6efc750f..16db933e0 100644--- a/funasr/models/qwen3_asr/model.py+++ b/funasr/models/qwen3_asr/model.py@@ -7,57 +7,51 @@ import torchimport torch.nn as nnfrom funasr.register import tables-from funasr.utils.load_utils import load_audio_text_image_video@tables.register("model_classes", "Qwen3ASR")class Qwen3ASR(nn.Module):- def __init__(- self,- model: str = "Qwen/Qwen3-ASR-1.7B",- model_conf: dict = None,- **kwargs,- ):+ def __init__(self, **kwargs):super().__init__()- model_conf = model_conf or {}- self.model_path = model- self.qwen3_asr_model = None- self._device = model_conf.get("device", kwargs.get("device", "cuda:0"))- self._dtype = model_conf.get("dtype", kwargs.get("dtype", "bf16"))- self._max_new_tokens = model_conf.get("max_new_tokens", 512)- self._max_inference_batch_size = model_conf.get("max_inference_batch_size", 32)- self._forced_aligner = model_conf.get("forced_aligner", None)- self._forced_aligner_kwargs = model_conf.get("forced_aligner_kwargs", None)+ model_path = kwargs.get("model_path", "Qwen/Qwen3-ASR-1.7B")+ device = kwargs.get("device", "cuda:0")+ dtype = kwargs.get("dtype", "bf16")+ max_new_tokens = kwargs.get("max_new_tokens", 512)+ max_inference_batch_size = kwargs.get("max_inference_batch_size", 32)+ forced_aligner = kwargs.get("forced_aligner", None)+ forced_aligner_kwargs = kwargs.get("forced_aligner_kwargs", None)+self._dtype_map = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}- self._init_model()+ self._device = device+ self.model_path = model_path++ self._placeholder = nn.Parameter(torch.empty(0))- def _init_model(self):try:from qwen_asr import Qwen3ASRModelexcept ImportError:raise ImportError(- "qwen-asr package is required for Qwen3ASR model. "- "Install with: pip install qwen-asr"+ "qwen-asr package is required. Install with: pip install qwen-asr")- dtype = self._dtype_map.get(self._dtype, torch.bfloat16)+ torch_dtype = self._dtype_map.get(dtype, torch.bfloat16)fa_kwargs = None- if self._forced_aligner:- fa_kwargs = self._forced_aligner_kwargs or {}- fa_kwargs.setdefault("dtype", dtype)- fa_kwargs.setdefault("device_map", self._device)+ if forced_aligner:+ fa_kwargs = forced_aligner_kwargs or {}+ fa_kwargs.setdefault("dtype", torch_dtype)+ fa_kwargs.setdefault("device_map", device)self.qwen3_asr_model = Qwen3ASRModel.from_pretrained(- self.model_path,- dtype=dtype,- device_map=self._device,- forced_aligner=self._forced_aligner,+ model_path,+ dtype=torch_dtype,+ device_map=device,+ forced_aligner=forced_aligner,forced_aligner_kwargs=fa_kwargs,- max_inference_batch_size=self._max_inference_batch_size,- max_new_tokens=self._max_new_tokens,+ max_inference_batch_size=max_inference_batch_size,+ max_new_tokens=max_new_tokens,)- logging.info(f"Qwen3ASR model loaded from {self.model_path}")+ logging.info(f"Qwen3ASR model loaded from {model_path}")def forward(self, **kwargs):raise NotImplementedError("Qwen3ASR only supports inference mode")@@ -71,44 +65,43 @@ class Qwen3ASR(nn.Module):frontend=None,**kwargs,):+ meta_data = {}+ time1 = time.perf_counter()+language = kwargs.get("language", None)return_time_stamps = kwargs.get("return_time_stamps", False)context = kwargs.get("context", "")- fs = kwargs.get("fs", 16000)-- meta_data = {}- time1 = time.perf_counter()if isinstance(data_in, (list, tuple)):- audio_list = data_in+ audio_inputs = list(data_in)elif isinstance(data_in, str):- audio_list = [data_in]+ audio_inputs = [data_in]elif isinstance(data_in, torch.Tensor):- audio_np = data_in.cpu().numpy()+ audio_np = data_in.cpu().numpy().astype(np.float32)if audio_np.ndim == 1:- audio_list = [(audio_np, fs)]+ audio_inputs = [(audio_np, 16000)]else:- audio_list = [(audio_np[i], fs) for i in range(audio_np.shape[0])]+ audio_inputs = [(audio_np[i], 16000) for i in range(audio_np.shape[0])]elif isinstance(data_in, np.ndarray):if data_in.ndim == 1:- audio_list = [(data_in, fs)]+ audio_inputs = [(data_in.astype(np.float32), 16000)]else:- audio_list = [(data_in[i], fs) for i in range(data_in.shape[0])]+ audio_inputs = [(data_in[i].astype(np.float32), 16000) for i in range(data_in.shape[0])]else:- audio_list = [data_in]+ audio_inputs = [data_in]time2 = time.perf_counter()meta_data["load_data"] = f"{time2 - time1:0.3f}"results = self.qwen3_asr_model.transcribe(- audio=audio_list,+ audio=audio_inputs,context=context,language=language,return_time_stamps=return_time_stamps,)time3 = time.perf_counter()- meta_data["inference"] = f"{time3 - time2:0.3f}"+ meta_data["batch_data_time"] = time3 - time2output = []for i, r in enumerate(results):diff --git a/tests_models/test_qwen3_asr.py b/tests_models/test_qwen3_asr.pyindex 5232ec79b..1934294a4 100644--- a/tests_models/test_qwen3_asr.py+++ b/tests_models/test_qwen3_asr.py@@ -1,5 +1,5 @@#!/usr/bin/env python3-"""Test Qwen3-ASR: multi-language speech recognition"""+"""Test Qwen3-ASR: multi-language speech recognition via AutoModel"""import sysimport time@@ -7,15 +7,16 @@ import timedef main():from funasr import AutoModel- model_id = "Qwen/Qwen3-ASR-1.7B"+ model_path = "Qwen/Qwen3-ASR-1.7B"url_zh = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_zh.wav"url_en = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav"print("[Qwen3-ASR] Loading model...")t0 = time.time()model = AutoModel(- model=model_id,+ model="Qwen3ASR",model_conf={+ "model_path": model_path,"dtype": "bf16","max_new_tokens": 256,"max_inference_batch_size": 32,@@ -25,7 +26,7 @@ def main():)print("[Qwen3-ASR] Model loaded in %.1fs" % (time.time() - t0))- # Test 1: Chinese audio+ # Test 1: Chinese audio with forced languageprint("[Qwen3-ASR] Test 1: Chinese inference...")t0 = time.time()res = model.generate(@@ -41,7 +42,7 @@ def main():print("[Qwen3-ASR] Test 1 FAILED - no text in result")return 1- # Test 2: English audio+ # Test 2: English audio with forced languageprint("[Qwen3-ASR] Test 2: English inference...")t0 = time.time()res = model.generate(