Librosa简介

Librosa是一个 Python 模块，用于分析一般的音频信号，是一个非常强大的python语音信号处理的第三方库，根据网络资料以及官方教程，本文主要总结了一些重要且常用的功能。

# 安装
!pip install librosa

Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
Requirement already satisfied: librosa in f:\programdata\anaconda3\envs\tf\lib\site-packages (0.8.1)
Requirement already satisfied: scikit-learn!=0.19.0,>=0.14.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (0.24.2)
Requirement already satisfied: joblib>=0.14 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (1.1.0)
Requirement already satisfied: packaging>=20.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (21.3)
Requirement already satisfied: numpy>=1.15.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (1.19.5)
Requirement already satisfied: pooch>=1.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (1.5.2)
Requirement already satisfied: decorator>=3.0.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (4.4.2)
Requirement already satisfied: audioread>=2.0.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (2.1.9)
Requirement already satisfied: resampy>=0.2.2 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (0.2.2)
Requirement already satisfied: scipy>=1.0.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (1.5.4)
Requirement already satisfied: soundfile>=0.10.2 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (0.10.3.post1)
Requirement already satisfied: numba>=0.43.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from librosa) (0.53.1)
Requirement already satisfied: setuptools in f:\programdata\anaconda3\envs\tf\lib\site-packages (from numba>=0.43.0->librosa) (58.0.4)
Requirement already satisfied: llvmlite<0.37,>=0.36.0rc1 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from numba>=0.43.0->librosa) (0.36.0)
Requirement already satisfied: pyparsing!=3.0.5,>=2.0.2 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from packaging>=20.0->librosa) (3.0.6)
Requirement already satisfied: appdirs in f:\programdata\anaconda3\envs\tf\lib\site-packages (from pooch>=1.0->librosa) (1.4.4)
Requirement already satisfied: requests in f:\programdata\anaconda3\envs\tf\lib\site-packages (from pooch>=1.0->librosa) (2.26.0)
Requirement already satisfied: six>=1.3 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from resampy>=0.2.2->librosa) (1.15.0)
Requirement already satisfied: threadpoolctl>=2.0.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from scikit-learn!=0.19.0,>=0.14.0->librosa) (3.0.0)
Requirement already satisfied: cffi>=1.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from soundfile>=0.10.2->librosa) (1.15.0)
Requirement already satisfied: pycparser in f:\programdata\anaconda3\envs\tf\lib\site-packages (from cffi>=1.0->soundfile>=0.10.2->librosa) (2.21)
Requirement already satisfied: certifi>=2017.4.17 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from requests->pooch>=1.0->librosa) (2021.5.30)
Requirement already satisfied: charset-normalizer~=2.0.0 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from requests->pooch>=1.0->librosa) (2.0.7)
Requirement already satisfied: idna<4,>=2.5 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from requests->pooch>=1.0->librosa) (3.3)
Requirement already satisfied: urllib3<1.27,>=1.21.1 in f:\programdata\anaconda3\envs\tf\lib\site-packages (from requests->pooch>=1.0->librosa) (1.26.7)

import numpy as np
import pandas as pd
import osimport IPython.display as ipd

加载音频文件

import librosa
audio_data = 'data/Data_MGTV/angry/audio_1027.wav'
x , sr = librosa.load(audio_data)
print(x.shape, sr)

(45159,) 22050

x

array([-0.11979394, -0.10811259, -0.04991762, ...,  0.00769441,0.00752225,  0.        ], dtype=float32)

print('x:', x, '\n')
print('x shape:', np.shape(x), '\n')
print('Sample Rate (KHz):', sr, '\n')
print('Check Len of Audio:', np.shape(x)[0]/sr)

x: [-0.11979394 -0.10811259 -0.04991762 ...  0.00769441  0.007522250.        ] x shape: (45159,) Sample Rate (KHz): 22050 Check Len of Audio: 2.048027210884354

读取时长

d = librosa.get_duration(y=x, sr=22050, S=None, n_fft=2048, hop_length=512, center=True, filename=None)
d

2.048027210884354

采样率

sr = librosa.get_samplerate(audio_data)
sr

16000

去除两端沉默

audio_file, _ = librosa.effects.trim(x)
print('Audio File:', audio_file, '\n')
print('Audio File shape:', np.shape(audio_file))

Audio File: [-0.11979394 -0.10811259 -0.04991762 ...  0.00769441  0.007522250.        ] Audio File shape: (45159,)

播放音频

IPython.display.Audio 可以让我们直接在 jupyter notebook 中播放音频，比如下面包房一段音频

ipd.Audio(audio_data)

波形图

在这里，我们绘制了一个简单的音频波形图。波图让我们知道给定时间的音频响度。

%matplotlib inline

import sklearn
import matplotlib.pyplot as plt
import librosa.displayplt.figure(figsize=(20, 5))
librosa.display.waveplot(y, sr=sr)
plt.show()

Spectogram

频谱图（Spectogram）是声音频率随时间变化的频谱的可视化表示，是给定音频信号的频率随时间变化的表示。'.stft' 将数据转换为短期傅里叶变换。STFT转换信号，以便我们可以知道给定时间给定频率的幅度。使用 STFT，我们可以确定音频信号在给定时间播放的各种频率的幅度。

Spectrogram特征是目前在语音识别和环境声音识别中很常用的一个特征，由于CNN在处理图像上展现了强大的能力，使得音频信号的频谱图特征的使用愈加广泛，甚至比MFCC使用的更多。

X = librosa.stft(x)
Xdb = librosa.amplitude_to_db(abs(X))
plt.figure(figsize=(20, 5))
librosa.display.specshow(Xdb, sr=sr, x_axis='time', y_axis='hz')
plt.colorbar()
plt.show()

librosa.display.specshow(Xdb, sr=sr, x_axis='time', y_axis='log')
plt.colorbar()

<matplotlib.colorbar.Colorbar at 0x24f53d3e6d8>

梅尔频率倒谱系数（MFCC）

信号的梅尔频率倒谱系数 (MFCC) 是一小组特征（通常约为 10-20），它们简明地描述了频谱包络的整体形状。在 MIR 中，它经常被用来描述音色。

#mcc
mfccs = librosa.feature.mfcc(y=x, sr=sr)

mfccs

array([[-1.47507828e+02, -1.39587173e+02, -1.63085953e+02, ...,-3.51147095e+02, -3.62041565e+02, -3.64722260e+02],[ 1.39314545e+02,  1.28688156e+02,  1.26540642e+02, ...,1.31368317e+02,  1.23287079e+02,  1.06071014e+02],[-5.88899651e+01, -7.76861572e+01, -8.52756119e+01, ...,-3.08440018e+01, -3.50476532e+01, -3.22384949e+01],...,[ 1.24901953e+01,  2.48859482e+01,  3.59340363e+01, ...,-3.30873656e+00, -5.68462515e+00, -5.88594961e+00],[-6.10755301e+00, -8.72181129e+00, -3.69202137e+00, ...,-2.46745777e+00, -7.76338100e+00, -8.60360718e+00],[-1.22752495e+01, -8.53678513e+00, -2.76085877e+00, ...,6.47896719e+00,  9.00872326e+00, -3.04730564e-01]], dtype=float32)

mfccs.shape

(20, 89)

在这个例子中，mfcc 在 89 帧中计算了 20 个 MFCC。

第一个 MFCC，第 0 个系数，不传达与频谱整体形状相关的信息。它只传达一个恒定的偏移量，即向整个频谱添加一个恒定值。因此，很多情况我们可以在进行分类时会丢弃第一个MFCC。

librosa.display.specshow(mfccs, sr=sr, x_axis='time')

<matplotlib.collections.QuadMesh at 0x24f53d58a20>

过零率

过零率（zero-crossing rate，ZCR）是指一个信号的符号变化的比率，例如信号从正数变成负数，或反过来。这个特征已在语音识别和音乐信息检索领域得到广泛使用，是分类敲击声的关键特征。为真时为1，否则为0。在一些应用场景下，只统计“正向”或“负向”的变化，而不是所有的方向。

n0 = 7000
n1 = 7025
plt.figure(figsize=(14, 5))
plt.plot(x[n0:n1])
plt.show()

zero_crossings = librosa.zero_crossings(x[n0:n1], pad=False)

zero_crossings.shape

(25,)

zero_crossings.sum()

2

可以使用整个音频来遍历这个并推断出整个数据的过零。

zcrs = librosa.feature.zero_crossing_rate(x)
print(zcrs.shape)

(1, 89)

plt.figure(figsize=(14, 5))
plt.plot(zcrs[0])

[<matplotlib.lines.Line2D at 0x24f4cd88eb8>]

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-pakxb1Os-1651758177390)(https://mmbiz.qlogo.cn/mmbiz_png/1FD1x61uYVdHWBWhQKLKptH3N9aicoJkXTUxeNaFAgV1cWuKOZpkrWPNQYgZk6ibOXm0AE60DkOvCibLtEI0by0cQ/0)]

频谱质心：Spectral Centroid

频谱质心（维基百科）表示频谱能量集中在哪个频率上。这就像一个加权平均值：

其中 S(k) 是频段 k 处的频谱幅度，f(k) 是频段 k 处的频率。

spectral_centroids = librosa.feature.spectral_centroid(x, sr=sr)[0]
spectral_centroids.shape

(89,)

frames = range(len(spectral_centroids))
t = librosa.frames_to_time(frames)

import sklearn
def normalize(x, axis=0):return sklearn.preprocessing.minmax_scale(x, axis=axis)

librosa.display.waveplot(x, sr=sr, alpha=0.4)
plt.plot(t, normalize(spectral_centroids), color='r')

[<matplotlib.lines.Line2D at 0x24f505d4320>]

频谱带宽:Spectral Bandwidth

librosa.feature.spectral_bandwidth 可以用来计算p-order频谱带宽：

其中 S(k) 是频段 k 处的频谱幅度，f(k) 是频段 k 处的频率，fc 是频谱质心。当 p=2 时，这就像一个加权标准差。

spectral_bandwidth_2 = librosa.feature.spectral_bandwidth(x+0.01, sr=sr)[0]
spectral_bandwidth_3 = librosa.feature.spectral_bandwidth(x+0.01, sr=sr, p=3)[0]
spectral_bandwidth_4 = librosa.feature.spectral_bandwidth(x+0.01, sr=sr, p=4)[0]
librosa.display.waveplot(x, sr=sr, alpha=0.4)
plt.plot(t, normalize(spectral_bandwidth_2), color='r')
plt.plot(t, normalize(spectral_bandwidth_3), color='g')
plt.plot(t, normalize(spectral_bandwidth_4), color='y')
plt.legend(('p = 2', 'p = 3', 'p = 4'))

<matplotlib.legend.Legend at 0x24f50c08470>

频谱滚降

频谱衰减是总频谱能量的特定百分比所在的频率。

spectral_rolloff = librosa.feature.spectral_rolloff(x+0.01, sr=sr)[0]
librosa.display.waveplot(x, sr=sr, alpha=0.4)
plt.plot(t, normalize(spectral_rolloff), color='r')

[<matplotlib.lines.Line2D at 0x24f4cc06e10>]

色度特征：Chroma Feature

色度向量 (Wikipedia) 是一个典型的 12 元素特征向量，指示每个音高类别{C, C#, D, D#, E, ..., B}的能量是多少存在于信号中。

chromagram = librosa.feature.chroma_stft(x, sr=sr, hop_length=512)
plt.figure(figsize=(15, 5))
librosa.display.specshow(chromagram, x_axis='time', y_axis='chroma', hop_length=512, cmap='coolwarm')

<matplotlib.collections.QuadMesh at 0x24f4cc9db70>

间距和幅度

音高是声音的感知属性，在与频率相关的尺度上排序，或者更常见的是，音高是可以判断声音在与音乐旋律相关的意义上“更高”和“更低”的质量。

pitches, magnitudes = librosa.piptrack(y=x, sr=sr)
print(pitches)

[[0. 0. 0. ... 0. 0. 0.][0. 0. 0. ... 0. 0. 0.][0. 0. 0. ... 0. 0. 0.]...[0. 0. 0. ... 0. 0. 0.][0. 0. 0. ... 0. 0. 0.][0. 0. 0. ... 0. 0. 0.]]

参考资料

• librosa语音信号处理

• 语音信号处理库 ——Librosa

往期精彩回顾适合初学者入门人工智能的路线及资料下载(图文+视频)机器学习入门系列下载中国大学慕课《机器学习》（黄海广主讲）机器学习及深度学习笔记等资料打印《统计学习方法》的代码复现专辑机器学习交流qq群955171419，加入微信群请扫码：