下面这段代码演示了如何使用官方 SDK 把音频文件送进服务，并提取文字结果。我们特别加入了分段处理逻辑，以免一次性喂入太长音频导致 OOM。

from qwen_asr import Qwen3ASRModel
from pydub import AudioSegment
import os, time
def split_audio_if_needed:
    """若音频超过 max_duration 秒则切割成若干段"""
    audio = AudioSegment.from_wav
    duration = len / 1000   # ms → s
    if duration <= max_duration:
        return , duration
    chunks = 
    num_chunks = int + 1
    for i in range:
        start = i * max_duration * 1000          # 转回 ms
        end   = min * max_duration * 1000, len)
        chunk = audio
        chunk_path = f"temp_chunk_{i}.wav"
        chunk.export
        chunks.append
    return chunks, duration
def extract_text:
    """统一返回字符串，无论是 list 还是单对象"""
    if isinstance:
        return " ".join])
    return getattr)
def main:
    model = Qwn​as​r​Model.from_pretrained(
        "./Qwen​_ASR‑1.​7B",      # 本地路径，请自行替换为实际目录 
        device_map='cpu',         # 若有 GPU 改为 'cuda'
        torch_dtype='auto'
    )
    audio_file = "sample.wav"
    if not os.path.exists:
        print
        return
    chunks, total_len = split_audio_if_needed
    all_texts = 
    start_all = time.time
    for idx, path in enumerate:
        try:
            res = model.transcribe
            txt = extract_text
            all_texts.append
            print}] ✅ 完成")
        except Exception as e:
            print
            all_texts.append
      elapsed = time.time - start_all 
      full_text = " ".join
      print
      print
if __name__ == "__main__":
      main

* 小技巧 *: 在循环里加上 Tqdm Ke以kan到geng友好的进度条；若服务器负载较高，可考虑把 bsize=8 参数去掉——它根本不被 API 支持。

model.to
model.half                # 将权重转为 float16 
torch.backends.cudnn.benchmark=True   # 启动 cudnn 自动调优