一.可视化
本节主要介绍如何将语音信号可视化，以及读取wav和label文件并保存为字典，列表。
输入：输入的wav文件所对应的数据矩阵wave_data和帧率framerate。
例：[[1507 1374 1218 ... -78 -127 -43]] ，16000
输出：可视化图，如波形图，频谱图
二.代码：
#coding=utf-8
import os
import wave
import numpy as np
import matplotlib.pyplot as plt
import math
import time
from python_speech_features import mfcc
from python_speech_features import delta
from python_speech_features import logfbank
from scipy.fftpack import fft
def read_wav_data(filename):
"""
:param filename:输入音频的绝对路径（路径+文件名）例：D:\GitHub\wav\dae\train\A2_1.wav
:return:wave_data,framerate:输出音频矩阵，帧率。例：A2_1 [[1507 1374 1218 ... -78 -127 -43]]
读取wav文件，返回声音信号的时域谱矩阵和播放时间
"""
wav = wave.open(filename,"rb") #打开wav格式的声音文件filename
audioname = filename.split('\')[-1] #音频名
num_frame = wav.getnframes() #获取帧数
#print("{}帧数为：{}".format(audioname,num_frame))
num_channel = wav.getnchannels() #获取声道数
#print("{}声道数为：{}".format(audioname,num_channel))
framerate = wav.getframerate() # 获取帧率
#print("{}帧率为：{}".format(audioname,framerate))
num_sample_width = wav.getsampwidth() #获取每一帧的比特宽度
#print("{}比特宽度为：{}".format(audioname,num_sample_width))
str_data = wav.readframes(num_frame) # 读取全部的帧(二进制字符串)
wav.close() # 关闭流
wave_data = np.fromstring(str_data,dtype=np.short) # 将声音文件数据从字符串格式转换为数组矩阵形式
# print("{} shape: {}".format("wave_data",wave_data.shape))
# print("{} type: {}".format("wave_data",wave_data.dtype))
# print("{}: {}".format("wave_data",wave_data))
wave_data.shape = -1, num_channel #按照声道数将数组整形，单声道是一列，双声道是两列矩阵
# print("{} shape(整形后): {}".format("wave_data",wave_data.shape))
# print("{}(整形后): {}".format("wave_data",wave_data))
wave_data = wave_data.T # 将wave_data矩阵转置
# print("{} shape(转置后)：{}".format("wave_data",wave_data.shape))
# print("{}(转置后)：{}".format("wave_data",wave_data))
# print("{} len:{}".format("wave_data[0]",len(wave_data[0])))
return wave_data, framerate
x = np.linspace(0, 400 - 1, 400, dtype = np.int64)
w = 0.54 - 0.46 * np.cos(2 * np.pi * (x) / (400 - 1) ) # 汉明窗
def GetFrequencyFeature(wavsignal, fs):
"""
:param wavsignal:音频矩阵 例：[[1507 1374 1218 ... -78 -127 -43]]
:param fs:帧率 例：16000
:return data_input：转成频域后的音频矩阵
"""
# wav波形 加时间窗以及时移10ms
time_window = 25 # 单位ms
window_length = fs / 1000 * time_window # 计算窗长度的公式，目前全部为400固定值
wav_arr = np.array(wavsignal) # wav_arr:[[1507 1374 1218 ... -78 -127 -43]]
wav_length = wav_arr.shape[1] # wav_arr.shape[0]:1,wav_arr.shape[1]:163000
range0_end = int(len(wavsignal[0]) / fs * 1000 - time_window) // 10 # 计算循环终止的位置，也就是最终生成的窗数
data_input = np.zeros((range0_end, 200), dtype=np.float) # 用于存放最终的频率特征数据
data_line = np.zeros((1, 400), dtype=np.float)
for i in range(0, range0_end):
p_start = i * 160 # 0,160,320,480
p_end = p_start + 400 # 400,560,720,880
data_line = wav_arr[0, p_start:p_end] # 分帧
data_line = data_line * w # 加窗(这里是汉明窗)
data_line = np.abs(fft(data_line)) / wav_length # 傅里叶变换
data_input[i] = data_line[0:200] # 设置为400除以2的值（即200）是取一半数据，因为是对称的
data_input = np.log(data_input + 1) # 取log
return data_input
def wav_scale(energy):
"""
:param energy:要进行归一化的语音信号
:return e:归一化后的语音信号
"""
means = energy.mean() # 均值
var=energy.var() # 方差
e=(energy-means)/math.sqrt(var) # 归一化能量
return e
def wav_show(wave_data, fs): # 显示出来声音波形
time = np.arange(0, len(wave_data)) * (1.0/fs) # 计算声音的播放时间，单位为秒
# 画声音波形
plt.subplot(111)
plt.plot(time, wave_data)
plt.xlabel('time/s')
plt.ylabel('value')
plt.show()
def get_wav_list(filename):
'''
读取一个wav文件列表，返回一个存储该列表的字典类型值
'''
txt_obj = open(filename, 'r') # 打开文件并读入
txt_text = txt_obj.read()
txt_lines = txt_text.split('
') # 文本分割
# print("txt_lines:
{}".format(txt_lines))
dic_filelist = {} # 初始化字典
list_wavmark = [] # 初始化wav列表
for i in txt_lines:
if (i != ''):
txt_l = i.split(' ')
dic_filelist[txt_l[0]] = txt_l[1]
list_wavmark.append(txt_l[0])
txt_obj.close()
# print("dic_filelist:
{}".format(dic_filelist))
# print("list_wavmark:
{}".format(list_wavmark))
return dic_filelist, list_wavmark
def get_wav_symbol(filename):
'''
读取指定数据集中，所有wav文件对应的语音符号
返回一个存储符号集的字典类型值
'''
txt_obj = open(filename, 'r') # 打开文件并读入
txt_text = txt_obj.read()
# print("txt_text:
{}".format(txt_text))
txt_lines = txt_text.split('
') # 文本分割
# print("txt_lines:
{}".format(txt_lines))
dic_symbol_list = {} # 初始化字典
list_symbolmark = [] # 初始化symbol列表
for i in txt_lines:
if (i != ''):
txt_l = i.split(' ')
dic_symbol_list[txt_l[0]] = txt_l[1:]
list_symbolmark.append(txt_l[0])
txt_obj.close()
# print("dic_symbol_list:
{}".format(dic_symbol_list))
print("list_symbolmark:
{}".format(list_symbolmark))
return dic_symbol_list, list_symbolmark
def GetSymbolList(datapath):
'''
加载拼音符号列表，用于标记符号
返回一个列表list类型变量
'''
txt_obj = open('dict.txt', 'r', encoding='UTF-8') # 打开文件并读入
txt_text = txt_obj.read()
#print("txt_text:
{}".format(txt_text))
txt_lines = txt_text.split('
') # 文本分割
#print("txt_lines:
{}".format(txt_lines))
list_symbol = [] # 初始化符号列表
for i in txt_lines:
if (i != ''):
txt_l = i.split(" ")
list_symbol.append(txt_l[0])
txt_obj.close()
list_symbol.append('_')
#print(list_symbol)
# SymbolNum = len(list_symbol)
return list_symbol
if(__name__=='__main__'):
wave_data, fs = read_wav_data("D:\GitHub\wav\dae\train\A2_1.wav")
#print("wave_data, fs:{}{}".format(wave_data, fs))
wav_show(wave_data[0], fs)
wave_scale = wav_scale(wave_data)
wav_show(wave_scale[0], fs)
t0 = time.time()
freimg = GetFrequencyFeature(wave_data, fs)
t1 = time.time()
print('time cost:', t1 - t0)
freimg = freimg.T
plt.subplot(111)
plt.imshow(freimg)
plt.colorbar(cax=None, ax=None, shrink=0.5)
plt.show()
# get_wav_list("D:\Code\pycharm\learning\20180903\train.wav.lst")
# get_wav_symbol("D:\Code\pycharm\learning\20180903\train.syllable.txt")
# GetSymbolList("D:\Code\pycharm\learning\20180910\")
三.程序输出：
