声音的本质是震动,震动的本质是位移关于时间的函数,波形文件(.wav)中记录了不同采样时刻的位移。
通过傅里叶变换,可以将时间域的声音函数分解为一系列不同频率的正弦函数的叠加,通过频率谱线的特殊分布,建立音频内容和文本的对应关系,以此作为模型训练的基础。
案例:画出语音信号的波形和频率分布
# -*- encoding:utf-8 -*-
import numpy as np
import numpy.fft as nf
import scipy.io.wavfile as wf
import matplotlib.pyplot as plt
sample_rate, sigs = wf.read('../machine_learning_date/freq.wav')
print(sample_rate) # 8000采样率
print(sigs.shape) # (3251,)
sigs = sigs / (2 ** 15) # 归一化
times = np.arange(len(sigs)) / sample_rate
freqs = nf.fftfreq(sigs.size, 1 / sample_rate)
ffts = nf.fft(sigs)
pows = np.abs(ffts)
plt.figure('Audio')
plt.subplot(121)
plt.title('Time Domain')
plt.xlabel('Time', fontsize=12)
plt.ylabel('Signal', fontsize=12)
plt.tick_params(labelsize=10)
plt.grid(linestyle=':')
plt.plot(times, sigs, c='dodgerblue', label='Signal')
plt.legend()
plt.subplot(122)
plt.title('Frequency Domain')
plt.xlabel('Frequency', fontsize=12)
plt.ylabel('Power', fontsize=12)
plt.tick_params(labelsize=10)
plt.grid(linestyle=':')
plt.plot(freqs[freqs >= 0], pows[freqs >= 0], c='orangered', label='Power')
plt.legend()
plt.tight_layout()
plt.show()
语音识别
梅尔频率倒谱系数(MFCC)通过与声音内容密切相关的13个特殊频率所对应的能量分布,可以使用梅尔频率倒谱系数矩阵作为语音识别的特征。基于隐马尔科夫模型进行模式识别,找到测试样本最匹配的声音模型,从而识别语音内容。
MFCC
梅尔频率倒谱系数相关API:
import scipy.io.wavfile as wf
import python_speech_features as sf
sample_rate, sigs = wf.read('../data/freq.wav')
mfcc = sf.mfcc(sigs, sample_rate)
案例:画出MFCC矩阵:
python -m pip install python_speech_features
import scipy.io.wavfile as wf
import python_speech_features as sf
import matplotlib.pyplot as mp
sample_rate, sigs = wf.read(
'../ml_data/speeches/training/banana/banana01.wav')
mfcc = sf.mfcc(sigs, sample_rate)
mp.matshow(mfcc.T, cmap='gist_rainbow')
mp.show()
隐马尔科夫模型
隐马尔科夫模型相关API:
import hmmlearn.hmm as hl
model = hl.GaussianHMM(n_components=4, covariance_type='diag', n_iter=1000)
# n_components: 用几个高斯分布函数拟合样本数据
# covariance_type: 相关矩阵的辅对角线进行相关性比较
# n_iter: 最大迭代上限
model.fit(mfccs) # 使用模型匹配测试mfcc矩阵的分值 score = model.score(test_mfccs)
案例:训练training文件夹下的音频,对testing文件夹下的音频文件做分类
语音识别设计思路
1、读取training文件夹中的训练音频样本,每个音频对应一个mfcc矩阵,每个mfcc都有一个类别(apple)
import os
import numpy as np
import scipy.io.wavfile as wf
import python_speech_features as sf
import hmmlearn.hmm as hl
# 1. 读取training文件夹中的训练音频样本,每个音频对应一个mfcc矩阵,每个mfcc都有一个类别(apple...)。
def search_file(directory):
"""
:param directory: 训练音频的路径
:return: 字典{'apple':[url, url, url ... ], 'banana':[...]}
"""
# 使传过来的directory匹配当前操作系统
directory = os.path.normpath(directory)
objects = {}
# curdir:当前目录
# subdirs: 当前目录下的所有子目录
# files: 当前目录下的所有文件名
for curdir, subdirs, files in os.walk(directory):
for file in files:
if file.endswith('.wav'):
label = curdir.split(os.path.sep)[-1] # os.path.sep为路径分隔符
if label not in objects:
objects[label] = []
# 把路径添加到label对应的列表中
path = os.path.join(curdir, file)
objects[label].append(path)
return objects
# 读取训练集数据
train_samples = search_file('../machine_learning_date/speeches/training')
2、把所有类别为apple的mfcc合并在一起,形成训练集。
训练集:
train_x:[mfcc1,mfcc2,mfcc3,...],[mfcc1,mfcc2,mfcc3,...]...
train_y:[apple],[banana]...
由上述训练集样本可以训练一个用于匹配apple的HMM。
#Python学习交流群:778463939
train_x, train_y = [], []
# 遍历字典
for label, filenames in train_samples.items():
# [('apple', ['url1,,url2...'])
# [("banana"),("url1,url2,url3...")]...
mfccs = np.array([])
for filename in filenames:
sample_rate, sigs = wf.read(filename)
mfcc = sf.mfcc(sigs, sample_rate)
if len(mfccs) == 0:
mfccs = mfcc
else:
mfccs = np.append(mfccs, mfcc, axis=0)
train_x.append(mfccs)
train_y.append(label)
3、训练7个HMM分别对应每个水果类别。 保存在列表中。
# 训练模型,有7个句子,创建了7个模型
models = {}
for mfccs, label in zip(train_x, train_y):
model = hl.GaussianHMM(n_components=4, covariance_type='diag', n_iter=1000)
models[label] = model.fit(mfccs) # # {'apple':object, 'banana':object ...}
4、读取testing文件夹中的测试样本,整理测试样本
测试集数据:
test_x: [mfcc1, mfcc2, mfcc3...]
test_y :[apple, banana, lime]
# 读取测试集数据
test_samples = search_file('../machine_learning_date/speeches/testing')
test_x, test_y = [], []
for label, filenames in test_samples.items():
mfccs = np.array([])
for filename in filenames:
sample_rate, sigs = wf.read(filename)
mfcc = sf.mfcc(sigs, sample_rate)
if len(mfccs) == 0:
mfccs = mfcc
else:
mfccs = np.append(mfccs, mfcc, axis=0)
test_x.append(mfccs)
test_y.append(label)
5、针对每一个测试样本:
1、分别使用7个HMM模型,对测试样本计算score得分。
2、取7个模型中得分最高的模型所属类别作为预测类别。
pred_test_y = []
for mfccs in test_x:
# 判断mfccs与哪一个HMM模型更加匹配
best_score, best_label = None, None
# 遍历7个模型
for label, model in models.items():
score = model.score(mfccs)
if (best_score is None) or (best_score < score):
best_score = score
best_label = label
pred_test_y.append(best_label)
print(test_y) # ['apple', 'banana', 'kiwi', 'lime', 'orange', 'peach', 'pineapple']
print(pred_test_y) # ['apple', 'banana', 'kiwi', 'lime', 'orange', 'peach', 'pineapple']
声音合成
根据需求获取某个声音的模型频域数据,根据业务需要可以修改模型数据,逆向生成时域数据,完成声音的合成。
案例,(数据集12.json地址):
import json
import numpy as np
import scipy.io.wavfile as wf
with open('../data/12.json', 'r') as f:
freqs = json.loads(f.read())
tones = [
('G5', 1.5),
('A5', 0.5),
('G5', 1.5),
('E5', 0.5),
('D5', 0.5),
('E5', 0.25),
('D5', 0.25),
('C5', 0.5),
('A4', 0.5),
('C5', 0.75)]
sample_rate = 44100
music = np.empty(shape=1)
for tone, duration in tones:
times = np.linspace(0, duration, duration * sample_rate)
sound = np.sin(2 * np.pi * freqs[tone] * times)
music = np.append(music, sound)
music *= 2 ** 15
music = music.astype(np.int16)
wf.write('../data/music.wav', sample_rate, music)