谱图：
注：FFT（快速傅里叶变换）；DFT（离散傅里叶变换）；DCT（离散余弦变换）；VMD（变分模态分解）
MEDC：张钰莎,蒋盛益.基于MFCC特征提取和改进SVM的语音情感数据挖掘分类识别方法研究[J].计算机应用与软件,2020,37(08):160-165+212.
VGFCC：刘雨柔,张雪英,陈桂军,黄丽霞,张静.VMD改进GFCC的情感语音特征提取[J].计算机工程与设计,2020,41(08):2265-2270.DOI:10.16208/j.issn1000-7024.2020.08.028.
音频文件读取：
import torchaudio
def open_audio(audio_file): # Load an audio file.
sig, sr = torchaudio.load(audio_file) # 加载音频文件
return sig, sr # 音频时间序列， 音频采样率
通道数调整：
import torch
def re_channel(aud, new_channel): # 统一音频通道数
sig, sr = aud # aud=（音频时间序列， 音频采样率）
if sig.shape[0] == new_channel: # 保持通道数
return aud # Nothing to do
if new_channel == 1: # 转换成单通道
resign = sig[:1, :] # Convert from stereo to mono by selecting only the first channel
else: # 转换成多通道
resign = torch.cat([sig, sig]) # Convert from mono to stereo by duplicating the first channel
return resign, sr
采样率调整：
import torch
import torchaudio
def resample(aud, new_sr): # 标准化采样率
sig, sr = aud
if sr == new_sr: # 若采样率与初始相同，则不需重新采样
return aud # Nothing to do
num_channels = sig.shape[0] # 获取原通道个数
resign = torchaudio.transforms.Resample(sr, new_sr)(sig[:1, :]) # Resample first channel
if num_channels > 1:
# Resample the second channel and merge both channels
re_two = torchaudio.transforms.Resample(sr, new_sr)(sig[1:, :])
resign = torch.cat([resign, re_two])
return resign, new_sr
音频长度调整：
import torchaudio
import random
def pad_trunc(aud, max_ms): # 调整为相同长度
sig, sr = aud
num_rows, sig_len = sig.shape
max_len = sr // 1000 * max_ms
if sig_len > max_len: # 裁剪多余部分
sig = sig[:, :max_len] # Truncate the signal to the given length
elif sig_len < max_len: # 全零补充
# Length of padding to add at the beginning and end of the signal
pad_begin_len = random.randint(0, max_len - sig_len)
pad_end_len = max_len - sig_len - pad_begin_len
pad_begin = torch.zeros((num_rows, pad_begin_len)) # Pad with 0s
pad_end = torch.zeros((num_rows, pad_end_len))
sig = torch.cat((pad_begin, sig, pad_end), 1) # inputs,dim=1
return sig, sr
获取谱图特征：
from torchaudio import transforms
def spectro_gram(aud, feature, n_mel=64, n_fft=1024, hop_len=None): # 谱图
sig, sr = aud
top_db = 80
feature_choice = feature # 谱图；Mel谱图；MFCC
# spec has shape [channel, n_mel, time], where channel is mono, stereo etc
# 音频信号的采样率，win_length窗口大小(默认：n_fft，FFT 的大小)，STFT 窗口之间的跳跃长度(默认：win_length // 2)，梅尔滤波器组的数量
# MelSpectrogram返回了一个函数名，故后面加了函数需要输入的值
if feature_choice == 'Spectrogram':
spec = transforms.Spectrogram(n_fft=n_fft, hop_length=hop_len)(sig)
# Convert to decibels
spec = transforms.AmplitudeToDB(top_db=top_db)(spec)
elif feature_choice == 'MelSpectrogram':
spec = transforms.MelSpectrogram(sr, n_fft=n_fft, hop_length=hop_len, n_mels=n_mel)(sig)
# Convert to decibels
spec = transforms.AmplitudeToDB(top_db=top_db)(spec)
else:
spec = transforms.MFCC(sr, melkwargs={"n_fft": n_fft, "hop_length": hop_len,
"n_mels": n_mel})(sig)
return spec
获取3D谱图特征（谱图及其一阶差分和二阶差分）：
import numpy as np
import torch
def delta_delta(spector, h):
right = np.concatenate([spector[:, 0].reshape((h, -1)), spector], axis=1)[:, :-1]
delta = (spector - right)[:, 1:]
delta_pad = delta[:, 0].reshape((h, -1))
delta = np.concatenate([delta_pad, delta], axis=1)
return delta
def get_3d_spec(spector, moments=None): #spector：谱图
if moments is not None:
(base_mean, base_std, delta_mean, delta_std,
delta2_mean, delta2_std) = moments
else:
base_mean, delta_mean, delta2_mean = (0, 0, 0)
base_std, delta_std, delta2_std = (1, 1, 1)
h, w = spector.shape
delta = delta_delta(spector, h)
delta2 = delta_delta(delta, h)
base = (spector - base_mean) / base_std
delta = (delta - delta_mean) / delta_std
delta2 = (delta2 - delta2_mean) / delta2_std
stacked = [arr.reshape((h, w, 1)) for arr in (base, delta, delta2)]
return torch.from_numpy(np.concatenate(stacked, axis=2))
注：一阶差分就是离散函数中连续相邻两项之差【定义X(k),则Y(k)=X(k+1)-X(k)就是此函数的一阶差分，即当前语音帧与前一帧之间的关系， 体现帧与帧（相邻两帧）之间的联系】
二阶差分表示的是一阶差分与一阶差分之间的关系【在一阶差分的基础上，Z(k)=Y(k+1)-Y(k)=X(k+2)-2*X(k+1)+X(k)为此函数的二阶差分，即前一阶差分与后一阶差分之间的关系，体现到帧上就是相邻三帧之间的动态关系】
绘制音频波形：
import matplotlib
matplotlib.use('TkAgg')
import matplotlib.pyplot as plt
import numpy as np
# 显示语音时域波形
time = np.arange(0, len(dur_aud[0][0])) * (1.0 / dur_aud[1]) #dur_aud=（音频时间序列， 音频采样率）
plt.plot(time, dur_aud[0][0])
plt.title("语音信号时域波形", fontproperties='Microsoft YaHei')
plt.xlabel("时长（秒）", fontproperties='SimHei')
plt.ylabel("振幅", fontproperties='SimHei')
plt.savefig("./img_data/语音信号时域波形图", dpi=600)
plt.show()
相关知识点：python实现语音信号的分帧、加窗、预处理_python_吃块小西瓜-DevPress官方社区 (csdn.net)
语音特征提取: 看懂梅尔语谱图(Mel-spectrogram)、梅尔倒频系数(MFCCs)的原理_melspectrogram_BeichenLiu.Polaris的博客-CSDN博客