Wenet--离线语音识别快速上手体验

操作系统：Linux
GPU与CUDA：NVIDIA GeForce RTX 4090 D，CUDA 12.5

1. 新建python环境

bash 复制代码

conda create -n wenet_offline python=3.10 -y
conda activate wenet_offline

2. 安装 Wenet

WeNet 提供了封装好的 Python 包，安装极其简单。这只是一个运行时库，包含了预训练模型和推理接口，非常适合快速使用。它会自动安装核心依赖，如 PyTorch。如果想进行模型训练或深度定制，则需要克隆其源码仓库，但对于"仅使用"来说，这就够了。

bash 复制代码

pip install git+https://gitcode.com/gh_mirrors/we/wenet

上面命令自动安装的 PyTorch /Numpy等版本兼容性有问题，重装下：

bash 复制代码

pip install torch==2.2.2+cu121 torchaudio==2.2.2+cu121 -f https://download.pytorch.org/whl/torch_stable.html

pip uninstall numpy -y
pip install "numpy<2"

3. 离线语音识别

python 复制代码

import wenet

model = wenet.load_model('paraformer')
result = model.transcribe('test_cn.wav')
print(result.text)

result = model.transcribe('test_en.wav')
print(result.text)

result = model.transcribe('test_mix.wav')
print(result.text)

python 复制代码

今天天气怎么样明天天气怎么样
what's your name hello who are you
hello你叫什么名字呀今天的天气怎么样ok bye

4. 多模型对比

python 复制代码

import wenet
import time

def test_models_compact():
    """紧凑格式测试模型"""
    
    audio_files = ['test_cn.wav', 'test_en.wav', 'test_mix.wav']
    models = ['paraformer', 'firered', 'whisper-large-v3', 'wenetspeech']
    
    print("模型对比测试 (格式: 模型-文件-时间-结果)")
    print("="*80)
    
    for model_name in models:
        try:
            model = wenet.load_model(model_name)
            
            for audio_file in audio_files:
                start = time.time()
                result = model.transcribe(audio_file)
                elapsed = time.time() - start
                
                result_text = result.text if hasattr(result, 'text') else str(result)
                # 单行输出
                print(f"{model_name:20} | {audio_file:15} | {elapsed:6.3f}s | {result_text}")
                
        except Exception as e:
            print(f"{model_name:20} | 失败: {e}")

if __name__ == "__main__":
    test_models_compact()

python 复制代码

模型对比测试 (格式: 模型-文件-时间-结果)
================================================================================
paraformer           | test_cn.wav     |  0.129s | 今天天气怎么样明天天气怎么样
paraformer           | test_en.wav     |  0.111s | what's your name hello who are you
paraformer           | test_mix.wav    |  0.222s | hello你叫什么名字呀今天的天气怎么样ok bye
firered              | test_cn.wav     |  0.913s | 今天天气怎么样明天天气怎么样
firered              | test_en.wav     |  0.784s | WHAT'S YOUR NAME HELLO WHO ARE YOU
firered              | test_mix.wav    |  1.747s | 哈喽你叫什么名字呀今天的天气怎么样 OK拜
whisper-large-v3     | test_cn.wav     |  2.057s |  How is the weather today? How is the weather tomorrow?
whisper-large-v3     | test_en.wav     |  1.454s |  What's your name? Hello, who are you?
whisper-large-v3     | test_mix.wav    |  2.786s |  Hello 你叫什么名字呀? 今天的天气怎么样? OK, bye.
wenetspeech          | test_cn.wav     |  0.067s | 今天天气怎么样明天天气怎么样
wenetspeech          | test_en.wav     |  0.066s | 我叫要耐姆你好胡油
wenetspeech          | test_mix.wav    |  0.130s | 你好你叫什么名字今天的天气怎么样好拜

Wenet--离线语音识别 快速上手体验

Wenet--离线语音识别快速上手体验