好像只有非流式推理才有 timestamp

example_qwen3_asr_vllm_streaming.py

@@ -21,8 +21,11 @@ Note:
     pip install qwen-asr[vllm]
 """
 
+# pip install -U flash-attn --no-build-isolation
+
 from vllm import LLM as vLLM
 from vllm import SamplingParams
+import torch
 
 import io
 import urllib.request
@@ -92,18 +95,36 @@ def run_streaming_case(asr: Qwen3ASRModel, wav16k: np.ndarray, step_ms: int) ->
 def main() -> None:
     # Streaming is vLLM-only and no forced aligner supported.
     asr = Qwen3ASRModel.LLM(
-        model=ASR_MODEL_PATH,
-        gpu_memory_utilization=0.6,
+        model="./Qwen3-ASR-1.7B",
+        dtype=torch.bfloat16,
+        gpu_memory_utilization=0.7,
         max_model_len=4096,
         max_num_seqs=1,
         enforce_eager=True,
         max_new_tokens=32,  # set a small value for streaming
+        forced_aligner="./Qwen3-ForcedAligner-0.6B",
+        forced_aligner_kwargs=dict(
+            dtype=torch.bfloat16,
+            device_map="cuda:0",
+            # attn_implementation="flash_attention_2",
+        )
     )
 
     wav_path = "./common_voice_en_444_16k_mono.wav"
     wav, sr = sf.read(wav_path, dtype="float32", always_2d=False)
     wav16k = _resample_to_16k(np.asarray(wav, dtype=np.float32), int(sr))
 
+    results = asr.transcribe(
+        audio=wav16k,
+        language="English", # can also be set to None for automatic language detection
+        return_time_stamps=True,
+    )
+
+    for r in results:
+        print(r.language, r.text, r.time_stamps[0])
+
+
+
     for step_ms in [500, 1000, 2000, 4000]:
         run_streaming_case(asr, wav16k, step_ms)
 

qwen_asr/inference/utils.py

@@ -26,6 +26,7 @@ import soundfile as sf
 
 AudioLike = Union[
     str,                      # wav path / URL / base64
+    np.ndarray,               # waveform assumed to be 16kHz
     Tuple[np.ndarray, int],   # (waveform, sr)
 ]
 MaybeList = Union[Any, List[Any]]
@@ -181,6 +182,7 @@ def normalize_audio_input(a: AudioLike) -> np.ndarray:
 
     Supported inputs:
         - str: local file path / https URL / base64 audio string
+        - np.ndarray: waveform (assumed to be 16kHz)
         - (np.ndarray, sr): waveform and sampling rate
 
     Returns:
@@ -189,6 +191,8 @@ def normalize_audio_input(a: AudioLike) -> np.ndarray:
     """
     if isinstance(a, str):
         audio, sr = load_audio_any(a)
+    elif isinstance(a, np.ndarray):
+        audio, sr = a, SAMPLE_RATE
     elif isinstance(a, tuple) and len(a) == 2 and isinstance(a[0], np.ndarray):
         audio, sr = a[0], int(a[1])
     else: