深度有趣 | 29 方言种类分类

栏目: Python · 发布时间: 6年前

内容简介:对于每一个MFCC特征都输出一个概率分布,然后结合CTC算法即可实现语音识别相比之下,语音分类要简单很多,因为对于整个MFCC特征序列只需要输出一个分类结果即可语音分类和语音识别的区别,可以类比一下文本分类和序列标注的区别

对于每一个MFCC特征都输出一个概率分布,然后结合CTC算法即可实现语音识别

相比之下,语音分类要简单很多,因为对于整个MFCC特征序列只需要输出一个分类结果即可

语音分类和语音识别的区别,可以类比一下文本分类和序列标注的区别

具体实现时,只需要稍微修改一下网络结构即可

数据

使用科大讯飞方言种类识别AI挑战赛提供的数据,challenge.xfyun.cn/,初赛提供了6种方言,复赛提供了10种方言

每种方言包括30个人每人200条共计6000条训练数据,以及10个人每人50条共计500条验证数据

数据以pcm格式提供,可以理解为wav文件去掉多余信息之后,仅保留语音数据的格式

实现

以下以长沙、南昌、上海三种方言数据为例,介绍如何实现语音分类

加载库

# -*- coding:utf-8 -*-

import numpy as np
import os
from matplotlib import pyplot as plt
from mpl_toolkits.axes_grid1 import make_axes_locatable
%matplotlib inline
from sklearn.utils import shuffle
import glob
import pickle
from tqdm import tqdm

from keras.models import Model
from keras.preprocessing.sequence import pad_sequences
from keras.layers import Input, Activation, Conv1D, Add, Multiply, BatchNormalization, GlobalMaxPooling1D, Dropout
from keras.optimizers import Adam
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau

from python_speech_features import mfcc
import librosa
from IPython.display import Audio
import wave
复制代码

加载pcm文件,共1W8条训练数据,1.5K条验证数据

train_files = glob.glob('data/*/train/*/*.pcm')
dev_files = glob.glob('data/*/dev/*/*/*.pcm')

print(len(train_files), len(dev_files), train_files[0])
复制代码

整理每条语音数据对应的分类标签

labels = {'train': [], 'dev': []}
    
for i in tqdm(range(len(train_files))):
    path = train_files[i]
    label = path.split('/')[1]
    labels['train'].append(label)
    
for i in tqdm(range(len(dev_files))):
    path = dev_files[i]
    label = path.split('/')[1]
    labels['dev'].append(label)

print(len(labels['train']), len(labels['dev']))
复制代码

定义处理语音、pcm转wav、可视化语音的三个函数,由于语音片段长短不一,所以去除少于1s的短片段,对于长片段则切分为不超过3s的片段

mfcc_dim = 13
sr = 16000
min_length = 1 * sr
slice_length = 3 * sr

def load_and_trim(path, sr=16000):
    audio = np.memmap(path, dtype='h', mode='r')
    audio = audio[2000:-2000]
    audio = audio.astype(np.float32)
    energy = librosa.feature.rmse(audio)
    frames = np.nonzero(energy >= np.max(energy) / 5)
    indices = librosa.core.frames_to_samples(frames)[1]
    audio = audio[indices[0]:indices[-1]] if indices.size else audio[0:0]
    
    slices = []
    for i in range(0, audio.shape[0], slice_length):
        s = audio[i: i + slice_length]
        if s.shape[0] >= min_length:
            slices.append(s)
    
    return audio, slices

def pcm2wav(pcm_path, wav_path, channels=1, bits=16, sample_rate=sr):
    data = open(pcm_path, 'rb').read()
    fw = wave.open(wav_path, 'wb')
    fw.setnchannels(channels)
    fw.setsampwidth(bits // 8)
    fw.setframerate(sample_rate)
    fw.writeframes(data)
    fw.close()

def visualize(index, source='train'):
    if source == 'train':
        path = train_files[index]
    else:
        path = dev_files[index]
    print(path)
        
    audio, slices = load_and_trim(path)
    print('Duration: %.2f s' % (audio.shape[0] / sr))
    plt.figure(figsize=(12, 3))
    plt.plot(np.arange(len(audio)), audio)
    plt.title('Raw Audio Signal')
    plt.xlabel('Time')
    plt.ylabel('Audio Amplitude')
    plt.show()
    
    feature = mfcc(audio, sr, numcep=mfcc_dim)
    print('Shape of MFCC:', feature.shape)
    fig = plt.figure(figsize=(12, 5))
    ax = fig.add_subplot(111)
    im = ax.imshow(feature, cmap=plt.cm.jet, aspect='auto')
    plt.title('Normalized MFCC')
    plt.ylabel('Time')
    plt.xlabel('MFCC Coefficient')
    plt.colorbar(im, cax=make_axes_locatable(ax).append_axes('right', size='5%', pad=0.05))
    ax.set_xticks(np.arange(0, 13, 2), minor=False);
    plt.show()
    
    wav_path = 'example.wav'
    pcm2wav(path, wav_path)
    
    return wav_path

Audio(visualize(2))
复制代码

一句长沙话对应的波形和MFCC特征

深度有趣 | 29 方言种类分类

整理数据,查看语音片段的长度分布,最后得到了18890个训练片段,1632个验证片段

X_train = []
X_dev = []
Y_train = []
Y_dev = []
lengths = []

for i in tqdm(range(len(train_files))):
    path = train_files[i]
    audio, slices = load_and_trim(path)
    lengths.append(audio.shape[0] / sr)
    for s in slices:
        X_train.append(mfcc(s, sr, numcep=mfcc_dim))
        Y_train.append(labels['train'][i])

for i in tqdm(range(len(dev_files))):
    path = dev_files[i]
    audio, slices = load_and_trim(path)
    lengths.append(audio.shape[0] / sr)
    for s in slices:
        X_dev.append(mfcc(s, sr, numcep=mfcc_dim))
        Y_dev.append(labels['dev'][i])
    
print(len(X_train), len(X_dev))
plt.hist(lengths, bins=100)
plt.show()
复制代码
深度有趣 | 29 方言种类分类

将MFCC特征进行归一化

samples = np.vstack(X_train)
mfcc_mean = np.mean(samples, axis=0)
mfcc_std = np.std(samples, axis=0)
print(mfcc_mean)
print(mfcc_std)

X_train = [(x - mfcc_mean) / (mfcc_std + 1e-14) for x in X_train]
X_dev = [(x - mfcc_mean) / (mfcc_std + 1e-14) for x in X_dev]

maxlen = np.max([x.shape[0] for x in X_train + X_dev])
X_train = pad_sequences(X_train, maxlen, 'float32', padding='post', value=0.0)
X_dev = pad_sequences(X_dev, maxlen, 'float32', padding='post', value=0.0)
print(X_train.shape, X_dev.shape)
复制代码

对分类标签进行处理

from sklearn.preprocessing import LabelEncoder
from keras.utils import to_categorical

le = LabelEncoder()
Y_train = le.fit_transform(Y_train)
Y_dev = le.transform(Y_dev)
print(le.classes_)

class2id = {c: i for i, c in enumerate(le.classes_)}
id2class = {i: c for i, c in enumerate(le.classes_)}

num_class = len(le.classes_)
Y_train = to_categorical(Y_train, num_class)
Y_dev = to_categorical(Y_dev, num_class)
print(Y_train.shape, Y_dev.shape)
复制代码

定义产生批数据的迭代器

batch_size = 16

def batch_generator(x, y, batch_size=batch_size): 
    offset = 0
    while True:
        offset += batch_size
        
        if offset == batch_size or offset >= len(x):
            x, y = shuffle(x, y)
            offset = batch_size
            
        X_batch = x[offset - batch_size: offset]    
        Y_batch = y[offset - batch_size: offset]
        
        yield (X_batch, Y_batch)
复制代码

定义模型并训练,通过GlobalMaxPooling1D对整个序列的输出进行降维,从而变成标准的分类任务

epochs = 10
num_blocks = 3
filters = 128
drop_rate = 0.25

X = Input(shape=(None, mfcc_dim,), dtype='float32')

def conv1d(inputs, filters, kernel_size, dilation_rate):
    return Conv1D(filters=filters, kernel_size=kernel_size, strides=1, padding='causal', activation=None, dilation_rate=dilation_rate)(inputs)

def batchnorm(inputs):
    return BatchNormalization()(inputs)

def activation(inputs, activation):
    return Activation(activation)(inputs)

def res_block(inputs, filters, kernel_size, dilation_rate):
    hf = activation(batchnorm(conv1d(inputs, filters, kernel_size, dilation_rate)), 'tanh')
    hg = activation(batchnorm(conv1d(inputs, filters, kernel_size, dilation_rate)), 'sigmoid')
    h0 = Multiply()([hf, hg])
    
    ha = activation(batchnorm(conv1d(h0, filters, 1, 1)), 'tanh')
    hs = activation(batchnorm(conv1d(h0, filters, 1, 1)), 'tanh')
    
    return Add()([ha, inputs]), hs

h0 = activation(batchnorm(conv1d(X, filters, 1, 1)), 'tanh')
shortcut = []
for i in range(num_blocks):
    for r in [1, 2, 4, 8, 16]:
        h0, s = res_block(h0, filters, 7, r)
        shortcut.append(s)

h1 = activation(Add()(shortcut), 'relu')
h1 = activation(batchnorm(conv1d(h1, filters, 1, 1)), 'relu') # batch_size, seq_len, filters
h1 = batchnorm(conv1d(h1, num_class, 1, 1)) # batch_size, seq_len, num_class
h1 = GlobalMaxPooling1D()(h1) # batch_size, num_class
Y = activation(h1, 'softmax')

optimizer = Adam(lr=0.01, clipnorm=5)
model = Model(inputs=X, outputs=Y)
model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])

checkpointer = ModelCheckpoint(filepath='fangyan.h5', verbose=0)
lr_decay = ReduceLROnPlateau(monitor='loss', factor=0.2, patience=1, min_lr=0.000)

history = model.fit_generator(
    generator=batch_generator(X_train, Y_train), 
    steps_per_epoch=len(X_train) // batch_size,
    epochs=epochs, 
    validation_data=batch_generator(X_dev, Y_dev), 
    validation_steps=len(X_dev) // batch_size, 
    callbacks=[checkpointer, lr_decay])
复制代码

绘制损失函数曲线和正确率曲线,经过10轮的训练后,训练集的正确率已经将近100%,而验证集则不太稳定,大概在89%左右

train_loss = history.history['loss']
valid_loss = history.history['val_loss']
plt.plot(train_loss, label='train')
plt.plot(valid_loss, label='valid')
plt.legend(loc='upper right')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.show()

train_acc = history.history['acc']
valid_acc = history.history['val_acc']
plt.plot(train_acc, label='train')
plt.plot(valid_acc, label='valid')
plt.legend(loc='upper right')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.show()
复制代码

验证集结果不够好的原因可能是训练数据不足,虽然一共有1W8条训练数据,但实际上只有90个说话人

如果说话人更多一些、声音更多样一些,模型应该能够学到各种方言所对应的更为通用的特征

深度有趣 | 29 方言种类分类
深度有趣 | 29 方言种类分类

保存分类和方言名称之间的映射,以便后续使用

with open('resources.pkl', 'wb') as fw:
    pickle.dump([class2id, id2class, mfcc_mean, mfcc_std], fw)
复制代码

在单机上加载训练好的模型,随机选择一条语音进行分类

# -*- coding:utf-8 -*-

import numpy as np
from keras.models import load_model
from keras.preprocessing.sequence import pad_sequences
import librosa
from python_speech_features import mfcc
import pickle
import wave
import glob

with open('resources.pkl', 'rb') as fr:
    [class2id, id2class, mfcc_mean, mfcc_std] = pickle.load(fr)

model = load_model('fangyan.h5')

paths = glob.glob('data/*/dev/*/*/*.pcm')
path = np.random.choice(paths, 1)[0]
label = path.split('/')[1]
print(label, path)

mfcc_dim = 13
sr = 16000
min_length = 1 * sr
slice_length = 3 * sr

def load_and_trim(path, sr=16000):
    audio = np.memmap(path, dtype='h', mode='r')
    audio = audio[2000:-2000]
    audio = audio.astype(np.float32)
    energy = librosa.feature.rmse(audio)
    frames = np.nonzero(energy >= np.max(energy) / 5)
    indices = librosa.core.frames_to_samples(frames)[1]
    audio = audio[indices[0]:indices[-1]] if indices.size else audio[0:0]
    
    slices = []
    for i in range(0, audio.shape[0], slice_length):
        s = audio[i: i + slice_length]
        slices.append(s)
    
    return audio, slices

audio, slices = load_and_trim(path)
X_data = [mfcc(s, sr, numcep=mfcc_dim) for s in slices]
X_data = [(x - mfcc_mean) / (mfcc_std + 1e-14) for x in X_data]
maxlen = np.max([x.shape[0] for x in X_data])
X_data = pad_sequences(X_data, maxlen, 'float32', padding='post', value=0.0)
print(X_data.shape)

prob = model.predict(X_data)
prob = np.mean(prob, axis=0)
pred = np.argmax(prob)
prob = prob[pred]
pred = id2class[pred]
print('True:', label)
print('Pred:', pred, 'Confidence:', prob)
复制代码

最后再提一下,既然是对三维tensor做分类,那么就和文本分类问题极其相似,所以也可以考虑使用BiLSTM之类的其他模型


以上所述就是小编给大家介绍的《深度有趣 | 29 方言种类分类》,希望对大家有所帮助,如果大家有任何疑问请给我留言,小编会及时回复大家的。在此也非常感谢大家对 码农网 的支持!

查看所有标签

猜你喜欢:

本站部分资源来源于网络,本站转载出于传递更多信息之目的,版权归原作者或者来源机构所有,如转载稿涉及版权问题,请联系我们

Java核心技术·卷 I(原书第10版)

Java核心技术·卷 I(原书第10版)

[美] 凯.S.霍斯特曼(Cay S. Horstmann) / 周立新 等 / 机械工业出版社 / 2016-9 / CNY 119.00

Java领域最有影响力和价值的著作之一,由拥有20多年教学与研究经验的资深Java技术专家撰写(获Jolt大奖),与《Java编程思想》齐名,10余年全球畅销不衰,广受好评。第10版根据Java SE 8全面更新,同时修正了第9版中的不足,系统全面讲解了Java语言的核 心概念、语法、重要特性和开发方法,包含大量案例,实践性强。 一直以来,《Java核心技术》都被认为是面向高级程序员的经典教......一起来看看 《Java核心技术·卷 I(原书第10版)》 这本书的介绍吧!

JSON 在线解析
JSON 在线解析

在线 JSON 格式化工具

图片转BASE64编码
图片转BASE64编码

在线图片转Base64编码工具

XML 在线格式化
XML 在线格式化

在线 XML 格式化压缩工具