项目简介
深度学习程序的边缘部署以性能绝佳的C++为主(⊙﹏⊙),但遇到项目开发周期短,则以功能优先,一些复杂的算法和处理用C++写怕不是得写到天荒地老,于是C++调用python以及第三方库的C端接口这样的方案就应运而生,牺牲一小部分性能,换来功能的完成,连准确性也顺便验证了(注意如果开发人员水平不够(ㄒoㄒ),用C++造轮子的性能还不如python)本项目首先开发了一个python的类用于预处理wav音频文件来提取MFCC特征,得益于python_speech_features库其实几行代码就能解决,但为了后续的学习借鉴,本次开发较完善点,开发的多个接口对多种数据传递的情况做演示,然后用C++调用这些python接口并取回数据,经测试,每次调用接口会比纯python执行慢不到1毫秒,最终打包后的项目放到无任何开发环境的虚拟机做测试,这其中的波折和踩坑真的只有做过的才懂┭┮﹏┭┮
梅尔频率倒谱系数(MFCC)通过对音频信号的处理和分析,提取出反映语音特征的信息,广泛应用于语音识别、语音合成、说话人识别等领域。可以简单的理解为将一个音频文件转为了矩阵,该矩阵保存了音频特征。
# 程序/数据集下载
本文章只发布于博客园和爆米算法,被抄袭后可能排版错乱或下载失效,作者:爆米LiuChen
环境准备
python3.8(虚拟环境或主环境均可)、VS2019(已支持cmake)、什么都没装的win虚拟机(用于测试)整个项目的文件结构如下
执行步骤
1.新建python虚拟环境
anaconda的命令是【conda create -n 环境名 python=3.8】,然后pip安装下numpy、scipy,python_speech这几个包2.虚拟环境运行下python代码
AudioPreprocess.py代码如下,主要实现了AudioPreprocess这个类,作用是将wav文件先采样成numpy矩阵,然后提取MFCC特征from python_speech_features import mfcc
import scipy.io.wavfile
from numpy.typing import NDArray
from typing import Tuple
import numpy as np
def yell():
print('''Congratulations,you import 【AudioPreprocess】 successfully!!!''')
class AudioPreprocess():
def __init__(self,numcep:int=13,keepSecs:int=8):
'''
预处理类
:param numcep: MFCC特征数(通道数)
:param keepSecs: 一个wav文件读取后保留的秒数,不够则补0
'''
self.numcep = numcep
self.keepSecs = keepSecs
def readWave(self,wavePath:str)->Tuple[int,NDArray[np.int16]]:
'''
读取一个wave文件
:param wavePath: wav文件路径
:return: 采样率,采样
'''
samplerate, samples = scipy.io.wavfile.read(wavePath)
return samplerate, samples
def samples2MFCC(self,samplerate:int, samples:NDArray[np.int16])->NDArray[np.float32]:
'''
一个wav的采样转MFCC特征
:param samplerate: 采样率
:param samples: 采样
:return: MFCC特征 size=(channel,feature)
'''
samples = samples if len(samples.shape) <= 1 else samples[:, 0]
samples = samples[:int(self.keepSecs*samplerate)]
samples = np.pad(samples, pad_width=(0, int(samplerate * self.keepSecs) - samples.shape[0]), mode='constant',constant_values=(0, 0))
mfccFeature = mfcc(samples, samplerate=samplerate,numcep=self.numcep)
mfccFeature = np.transpose(mfccFeature,axes=(1,0))
return mfccFeature
def wave2MFCC(self,wavePath:str)->NDArray[np.float64]:
'''
wav路径转MFCC
:param wavePath: wav文件路径
:return: MFCC特征 size=(channel,feature)
'''
samplerate, samples = self.readWave(wavePath)
mfccFeature = self.samples2MFCC(samplerate, samples)
return mfccFeature
if __name__ == "__main__":
import time
path = "test.wav"
audioPreprocess = AudioPreprocess()
samplerate, samples = audioPreprocess.readWave(path)
t1 = time.time()
for i in range(100):
mfccFeature = audioPreprocess.wave2MFCC(path)
t2 = time.time()
print((t2-t1)*1000)
3.迁移虚拟环境
可以将整个虚拟环境都转移到项目中,这样最稳,但文件也最多,我是主要复制了下面几个文件和文件夹,并删除了Lib/site-packages里一些用不到的库,结果还是得要250多M,numpy和scipy这俩库太大了...其实可以尝试一个个的删除,只要留下的文件能支撑你的项目就行,但我这边就懒得这么做了4.编写Cmakelists.txt
因为需要调用python解释器,并且也用到了numpy的C接口,所以要额外编写下这俩的配置,需要的文件都在我们的虚拟环境中cmake_minimum_required (VERSION 3.8)
project ("AudioPrepocess")
SET(CMAKE_BUILD_TYPE "Release")#Debug或Release模式
set(CMAKE_CXX_STANDARD 11)
add_compile_options("$<$<C_COMPILER_ID:MSVC>:/utf-8>")
add_compile_options("$<$<CXX_COMPILER_ID:MSVC>:/utf-8>")
#项目文件路径配置
set(CMAKE_BINARY_DIR "${CMAKE_SOURCE_DIR}/build")#项目源码构建路径
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${CMAKE_SOURCE_DIR}/bin")#存放可执行软件的目录;
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_SOURCE_DIR}/lib")#默认存放项目生成的静态库的文件夹位置;
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY "${CMAKE_SOURCE_DIR}/lib")#默认存放项目生成的动态库的文件夹位置;
include_directories(include)#头文件目录
aux_source_directory(source SRC_FILES)#源文件目录的所有文件
#调用python的设置
set(PYTHON_DIR "${CMAKE_SOURCE_DIR}/python38/env")
include_directories("${PYTHON_DIR}/include")#头文件目录
link_libraries("${PYTHON_DIR}/libs/python38.lib")
#调用numpy的设置
include_directories("${CMAKE_SOURCE_DIR}/python38/env/Lib/site-packages/numpy/core/include/numpy")#头文件目录
link_libraries("${CMAKE_SOURCE_DIR}/python38/env/Lib/site-packages/numpy/core/lib/npymath.lib")
#移动一些python的依赖
file(COPY "${CMAKE_SOURCE_DIR}/python38" DESTINATION "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
file(RENAME "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/python38/env/python38.dll" "${CMAKE_SOURCE_DIR}/bin/python38.dll")
add_executable(${PROJECT_NAME} main.cpp ${SRC_FILES} "source/AudioPreprocess.cpp")#构建可执行文件
5.编写C++代码
include/AudioPreprocess.h如下,声明一个对应python的AudioPreprocess类,成员函数也一致(可以不用这么对应,单纯写个函数去调执行py脚本里的AudioPreprocess类接口就行)反正最后是调用python代码,要不要对应不重要,但这个博客主要是演示的全面一点,注释也写得全一点#include <chrono>
#include <vector>
#include "Python.h"
#include "arrayobject.h"
long long getCurrentTimeMS();//获得当前时间戳 单位毫秒
int* initNumpy();//初始化numpy会有返回值 不能直接放在类的构造函数中,所以拿个形式函数包裹下
//包裹readWave的返回值
struct ResReadWave {
int sampleRate;
PyArrayObject* samples;
};
//调用python进行音频预处理类 可选择是否标准化数据 但需要传入标准化文件路径
class AudioPreprocess
{
public:
/// @brief 初始化python 初始化模块和导入的python类
/// @param scalerPath 标准化文件路径
/// @param numcep MFCC特征数(通道数)
/// @param keepSecs 一个wav文件读取后保留的秒数,不够则补0
AudioPreprocess(int numcep=13, int keepSecs=8);
/// @brief 读取wav文件,返回采样率和采样
/// @param wavePath
ResReadWave readWave(char* wavePath);
/// @brief 采样转MFCC特征 返回MFCC特征
/// @param samplerates
/// @param samples
/// @return MFCC特征 二维数组
PyArrayObject *samples2MFCC(int samplerates, PyArrayObject* samples);
/// @brief 读取wav文件,返回MFCC特征
/// @param wavePath
/// @return MFCC特征 二维数组
PyArrayObject* wave2MFCC(char* wavePath);
~AudioPreprocess();
private:
PyObject* pyModule;
PyObject* pyFunc;
PyObject* pyArgs;
PyObject* pyClass;
PyObject* pyClassObj;
//python预处理类中对应的函数、参数、返回值
PyObject* pyFuncReadWave;
PyObject* pyArgsReadWave;
PyObject* pyResReadWave;
PyObject* pyFuncSamples2MFCC;
PyObject* pyArgsSamples2MFCC;
PyObject* pyResSamples2MFCC;
PyObject* pyFuncWave2MFCC;
PyObject* pyArgsWave2MFCC;
PyObject* pyResWave2MFCC;
int numcep;
int keepSecs;
};
source/AudioPreprocess.cpp如下,实现C++和python互传一些基本类型以及numpy这种稍微复杂点的矩阵,注意python初始化的执行顺序,还有最好手动释放那些python对象,还有注意numpy的数据精度类型,不对齐是不会报错的 可以看出C++其实实例化了一个python解释器,然后在解释器里执行python代码,等于在python外套了一层,因此不管怎样都不可能比python还快,这种方式适合需要实现复杂算法且开发时间短的场景,毕竟谁愿意去看MFCC的公式呢...
#include "AudioPreprocess.h"
int* initNumpy() {
import_array();
}
long long getCurrentTimeMS() {
auto now = std::chrono::system_clock::now(); // 获取当前时间点
auto now_ms = std::chrono::time_point_cast<std::chrono::milliseconds>(now); // 转换为毫秒
auto epoch = now_ms.time_since_epoch(); // 计算自纪元以来的毫秒数
return epoch.count(); // 返回毫秒数
}
AudioPreprocess::AudioPreprocess(int numcep, int keepSecs):numcep(numcep), keepSecs(keepSecs){
//初始化python解释器
Py_SetPythonHome(L"python38/env");
Py_Initialize();
initNumpy();//初始化numpy,必须紧跟在python解释器初始化后面
PyRun_SimpleString("import sys;sys.path.append('./python38')");
this->pyModule = PyImport_ImportModule("AudioPreprocess");
this->pyFunc = PyObject_GetAttrString(this->pyModule, "yell");//yell这个函数的作用只是确认导入成功 顺便示范下怎么调用python函数
PyEval_CallObject(this->pyFunc, nullptr);
//实例化python的音频处理类
this->pyClass = PyObject_GetAttrString(this->pyModule, "AudioPreprocess");//获取AudioPreprocess这个类
this->pyArgs = Py_BuildValue("(i,i)", numcep, keepSecs);
this->pyClassObj = PyEval_CallObject(this->pyClass,this->pyArgs);
//初始化指针对应python的音频处理类成员函数、参数、返回值
this->pyFuncReadWave = PyObject_GetAttrString(this->pyClassObj, "readWave");
this->pyArgsReadWave = PyTuple_New(1);
this->pyResReadWave = PyTuple_New(2);
this->pyFuncSamples2MFCC = PyObject_GetAttrString(this->pyClassObj, "samples2MFCC");
this->pyArgsSamples2MFCC = PyTuple_New(2);
this->pyResSamples2MFCC = PyTuple_New(1);
this->pyFuncWave2MFCC = PyObject_GetAttrString(this->pyClassObj, "wave2MFCC");
this->pyArgsWave2MFCC = PyTuple_New(1);
this->pyResWave2MFCC = PyTuple_New(1);
}
ResReadWave AudioPreprocess::readWave(char* wavePath) {
//传入路径
PyTuple_SetItem(this->pyArgsReadWave,0,Py_BuildValue("s",wavePath));
this->pyResReadWave = PyEval_CallObject(this->pyFuncReadWave, this->pyArgsReadWave);
//返回值1 采样率
int sampleRate;
PyArg_Parse(PyTuple_GetItem(this->pyResReadWave, 0),"i", &sampleRate);
//返回值2 采样 numpy int16一维数组
PyArrayObject* samples = (PyArrayObject*)PyArray_FROM_OTF(PyTuple_GetItem(this->pyResReadWave, 1), NPY_INT16, NPY_IN_ARRAY);
ResReadWave result = {sampleRate,samples};
//打印下值,验证准确性 python输出的值为58
npy_intp indices[1] = {0}; // [0]的位置
int16_t value = *(int16_t*)PyArray_GetPtr(result.samples, indices);
printf("python输出数组[0,0] :58\nC++&python输出数组[0,0]:%d\n\n",value);
return result;
}
PyArrayObject* AudioPreprocess::samples2MFCC(int sampleRate, PyArrayObject* samples) {
//传入 采样率 采样二维数组
PyTuple_SetItem(this->pyArgsSamples2MFCC, 0, Py_BuildValue("i", sampleRate));
PyTuple_SetItem(this->pyArgsSamples2MFCC, 1, (PyObject*)samples);
this->pyResSamples2MFCC = PyEval_CallObject(this->pyFuncSamples2MFCC, this->pyArgsSamples2MFCC);
//返回值 采样二维数组
PyArrayObject* mfccFeature = (PyArrayObject*)PyArray_FROM_OTF(this->pyResSamples2MFCC, NPY_FLOAT64, NPY_IN_ARRAY);
//打印下值,验证准确性 python输出的值为11.31785676885986
npy_intp indices[2] = {0,0}; // [0,0]的位置
double_t value = *(double_t*)PyArray_GetPtr(mfccFeature, indices);
printf("python输出数组[0,0] :11.31785676885986\nC++&python输出数组[0,0]:%.14f\n",value);
return mfccFeature;
}
PyArrayObject* AudioPreprocess::wave2MFCC(char* wavePath) {
//传入路径
PyTuple_SetItem(this->pyArgsWave2MFCC, 0, Py_BuildValue("s", wavePath));
this->pyResWave2MFCC = PyEval_CallObject(this->pyFuncWave2MFCC, this->pyArgsWave2MFCC);
//返回值 采样二维数组
PyArrayObject* mfccFeature = (PyArrayObject*)PyArray_FROM_OTF(this->pyResWave2MFCC, NPY_FLOAT64, NPY_IN_ARRAY);
return mfccFeature;
}
AudioPreprocess::~AudioPreprocess() {
Py_CLEAR(pyModule);
Py_CLEAR(pyFunc);
Py_CLEAR(pyArgs);
Py_CLEAR(pyClass);
Py_CLEAR(pyClassObj);
Py_CLEAR(pyFuncReadWave);
Py_CLEAR(pyArgsReadWave);
Py_CLEAR(pyResReadWave);
Py_CLEAR(pyFuncSamples2MFCC);
Py_CLEAR(pyArgsSamples2MFCC);
Py_CLEAR(pyResSamples2MFCC);
Py_CLEAR(pyFuncWave2MFCC);
Py_CLEAR(pyArgsWave2MFCC);
Py_CLEAR(pyResWave2MFCC);
Py_Finalize();
}
main.cpp如下,验证下上文实现的方法,并于python做下对比验证,精度不一致问题是深度学习大忌,还有看看性能损失有多少,顺便做一下多线程实验,python内部的GIL锁会导致C++多线程崩溃,必须手动给python加锁
#include <iostream>
#include "AudioPreprocess.h"
#include <thread>
#include <mutex>
AudioPreprocess AP(13, 8);//初始化音频处理类 理论上只需要简单实现wave2MFCC函数,但我对应python的类都实现了,就当练习
void wave2MFCC_thread(char* wavePath) {
PyGILState_STATE state = PyGILState_Ensure();
AP.wave2MFCC("./python38/test.wav");
PyGILState_Release(state);
}
void main() {
ResReadWave resReadWave;//存储采样率和采用
PyArrayObject* mfccFeature;//存储MFCC特征
//resReadWave.samples只能在类内访问 不明原因 可能是因为python解释器在那个类中初始化的,可以想办法在类内转成C++ vector数组再访问
resReadWave = AP.readWave("./python38/test.wav");
mfccFeature = AP.samples2MFCC(resReadWave.sampleRate, resReadWave.samples);
mfccFeature = AP.wave2MFCC("./python38/test.wav");
//运行100次,计算时间 ,对比纯python的时间
long long t1 = getCurrentTimeMS();
for (int i = 1; i <= 100; ++i) {
mfccFeature = AP.wave2MFCC("./python38/test.wav");
}
long long t2 = getCurrentTimeMS();
printf("\npython运行100次函数时间 :930 ms\nC++&python运行100次函数时间:%d ms\n",t2-t1);
//多线程实验 如果没处理好 C++多线程会使python解释器崩溃
printf("\n多线程实验");
printf("\n多线程初始化:%d", PyEval_ThreadsInitialized());
printf("\n全局解释器锁GIL:%d\n", PyGILState_Check());
//PyEval_InitThreads();//开启多线程支持 3.8这个版本已经不需要手动调用这行代码来开启多线程支持
Py_BEGIN_ALLOW_THREADS;//暂时释放全局解释器锁GIL
char* wavePath = "./python38/test.wav";
std::thread t1(wave2MFCC_thread, wavePath);
std::thread t2(wave2MFCC_thread, wavePath);
t1.join();
t2.join();
Py_END_ALLOW_THREADS;//重新获取全局解释器锁
//Python的对象最好都自己手动销毁
Py_CLEAR(resReadWave.samples);
Py_CLEAR(mfccFeature);
system("pause");
}