目前最相似的方案是moviepy 库但是这只能提取音频并保存到本地
我确实能够先保存再用其他的音频库读取，但由于这会让运行效率低下python雪上加霜，再加上本人精神洁癖的缘故坚决不使用这种方法
所以，有没有什么方法能解决我的问题呢
input = videopath.mp4
# 然后output记录音频的波形数据。。。。
output = nparray #最好是这个
output = list # 这个也行
output = tensor # 这个也行
https://github.com/JyzjYzjyZ/Format-conversion-of-in-memory-audio
import soundfile,librosa,torch, copy
import numpy as np
import scipy.interpolate as interpolate
import torchaudio,wave,torchvision #build时祈福求平安
'''
author:goodjin5
date:2022.3.16 --> 2022.3.19
-v 1.0.2 --> 1.0.3
=====================================
'''
def resample(sound,orig_sr,target_sr=16000,resample_type:str=None):
'''
:param sound: numpy-float64 maybe can run in float 32
:param orig_sr: 原始的采样率
:param target_sr: 目标采样率，与原相同或为0时直接返回
:param resample_type: 采用类型，默认自动分配fast或者best,可以强制指定 使用 'best', 'kaiser_best','fast','kaiser_fast'
:return: 采样后的nparray
'''
if orig_sr == target_sr or target_sr==0:
return sound,orig_sr
if resample_type in ['best', 'kaiser_best','fast','kaiser_fast']:
restype = ['kaiser_best','kaiser_best','kaiser_fast','kaiser_fast'][['best', 'kaiser_best','fast', 'kaiser_fast'].index(resample_type)]
else:
restype = ('kaiser_fast' if target_sr <= orig_sr else 'kaiser_best')
if __name__=='__main__':
print(resample_type,restype)
sound = librosa.resample(y=sound,orig_sr=orig_sr,target_sr=target_sr,res_type=restype)
return sound,target_sr
def read(path:str,sr:int=16000,mono:bool=True):
'''
支持音视频格式，所以ffmpeg支持的格式都可以
:param path: 音频路径
:param sr: 采样率 为0时采用原始的采样率
:param mono: 启用单声道
:return: numpy-float64
'''
try:
sound,orig_sr = soundfile.read(path,dtype='float64')
sound=sound.T
except:
sound, orig_sr = librosa.load(path=path,dtype=np.float64,res_type='kaiser_fast')
# to mono,and resample sr
if mono:
sound = librosa.to_mono(sound)
# if sr == 0 or target_sr == orig_sr:
# # 无需采样
# return sound,orig_sr
# sound = librosa.resample(y=sound,orig_sr=orig_sr,target_sr=target_sr,
# res_type=('kaiser_fast' if target_sr <= orig_sr else 'kaiser_best'))
sound,s = resample(sound,orig_sr,sr)
return sound,s
def toWave(sound):
'''
:param sound: numpy-float64
:return: waveFile -bytes
'''
# 32767
return (sound * 32767).astype('int16').tobytes()
def toTorch(sound):
'''
:param sound: numpy-float64
:return: tensor-float32
'''
sound = sound.astype('float32')
t = torch.from_numpy(copy.deepcopy(sound))
# 扩展维度，维度转换
return t.unsqueeze(1).permute(1, 0)
def toNumpy_int16(sound):
return (sound * 32767).astype('int16')
def torchToNumpy_float(sound):
'''
:param sound: tensor
:return: numpy_float float是16还是32由输入的tensor决定
'''
# 将所有shape为1的维度降维
return sound.squeeze().numpy()
def resamplrBylen(orig_list:list,target_len:int):
'''
同于标准重采样，此函数将len(list1)=x 从采样为len(list2)=y；y为指定的值，list2为输出
:param orig_list: 是list,重采样的源列表：list1
:param target_len: 重采样的帧数：y
:return: 重采样后的数组:list2
'''
orig_list_len =len(orig_list)
k = target_len/orig_list_len
x = [x*k for x in range(orig_list_len)]
x[-1] = 3572740
if x[-1]!=target_len:
# 线性更改越界结尾
x1=x[-2];y1=orig_list[-2];x2=x[-1];y2=orig_list[-1]
y_resa = (y2 - y1) * (target_len - x1) / (x2 - x1) + y1
x[-1] = target_len;orig_list[-1]=y_resa
# 使用了线性的插值方法，也可以根据需要改成其他的。推荐是线性
f = interpolate.interp1d(x,orig_list,'linear')
del x
resample_list = f([x for x in range(target_len)])
return resample_list
if __name__=='__main__':
# exaxmple
# You can test performance and usability here
# change 'AUDIO_PATH.wav' or 'Video_path.*' all format[ffmpeg]
path = 'AUDIO_PATH.wav'
path = r'D:\setup\wav\130.wav'
import matplotlib.pyplot as plt
numpy_float64,sr = read(path,0,True)
# 把target设置的过低会出现不能除以0的错误
res_low,sr = resample(numpy_float64,sr,target_sr=100,resample_type='kaiser_best')
# resample(numpy_float64,sr,50000,'kaiser_best')
wave = toWave(numpy_float64)
tensor = toTorch(numpy_float64)
numpy_int16 = toNumpy_int16(numpy_float64)
resamplr_bylen = resamplrBylen(numpy_float64.tolist(),5000)
plt.figure(figsize=(1000, 50), dpi=80)
plt.subplot(111)
plt.plot(resamplr_bylen, color='red',linestyle="-",alpha=1)
# plt.plot(numpy_float64.tolist(), color='green',linestyle="-",alpha=1)
plt.plot(res_low.tolist(), color='blue',linestyle="-",alpha=1)
plt.text(5, 0.4, 'InAudio,max{0},min{1}\nred:By len\nblue:By sr'
.format(max(numpy_float64.tolist()),min(numpy_float64.tolist())))
plt.show()
print('Passing the test')
调ffmpeg输出wav流到 stdout然后直接读，你可能要自己实现一个文件类型的read()等方法
np或者 torchaudio 底层不是c++吗..跟python效率没什么关系吧？