基于CTC的语音识别系统训练

Posted sunhongwen

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了基于CTC的语音识别系统训练相关的知识,希望对你有一定的参考价值。

目录

最小语音识别模型

  • 输入
    • 音频特征向量,共计26维的mfcc特征向量
    • 为了使神经网络能够统一格式,将[batch_size, time_step, 26]中的time_step定为一个较大的值,如果实际时长不足,采用补零的方法补齐
  • 输出
    • 输出为一个[batch_size, words_length]的张量,方便训练将word_length也取一个较大的常量,不足的补零
  • 网络用keras进行训练,利用backend设置ctc。

模型训练

import os
import numpy as np
import scipy.io.wavfile as wav
from collections import Counter
from python_speech_features import mfcc
from keras.models import Model
from keras.layers import Dense, Dropout, Input
from keras.layers import Lambda, Activation
from keras.layers.merge import add, concatenate
from keras import backend as K
from keras.optimizers import SGD, Adadelta
from keras.layers.recurrent import GRU
from keras.preprocessing.sequence import pad_sequences

# 获取音频文件列表及音频id
def genwavlist(wavpath):
    wavfiles = {}
    fileids = []
    for (dirpath, dirnames, filenames) in os.walk(wavpath):
        for filename in filenames:
            if filename.endswith(‘.wav‘):
                filepath = os.sep.join([dirpath, filename])
                fileid = filename.strip(‘.wav‘)
                wavfiles[fileid] = filepath
                fileids.append(fileid)
    return wavfiles,fileids

# 计算mfcc,并将特征补零为[500,26]的shape
def compute_mfcc(file):
    fs, audio = wav.read(file)
    mfcc_feat = mfcc(audio, samplerate=fs, numcep=26)
    mfcc_feat = mfcc_feat[::3]
    mfcc_feat = np.transpose(mfcc_feat)  
    mfcc_feat = pad_sequences(mfcc_feat, maxlen=500, dtype=‘float‘, padding=‘post‘, truncating=‘post‘).T
    return mfcc_feat

# 生成拼音映射到符号的map
def gendict(textfile_path):
    dicts = []
    textfile = open(textfile_path,‘r+‘)
    for content in textfile.readlines():
        content = content.strip(‘
‘)
        content = content.split(‘ ‘,1)[1]
        content = content.split(‘ ‘)
        dicts += (word for word in content)
    counter = Counter(dicts)
    words = sorted(counter)
    wordsize = len(words)
    word2num = dict(zip(words, range(wordsize)))
    return word2num,len(word2num)

# 利用字典,将text映射为number
def text2num(textfile_path):
    lexcion,wordnum = gendict(textfile_path)
    word2num = lambda word:lexcion.get(word, 0)
    textfile = open(textfile_path, ‘r+‘)
    content_dict = {}
    for content in textfile.readlines():
        content = content.strip(‘
‘)
        cont_id = content.split(‘ ‘,1)[0]
        content = content.split(‘ ‘,1)[1]
        content = content.split(‘ ‘)
        content = list(map(word2num,content))
        add_num = list(np.zeros(50-len(content)))
        content = content + add_num
        content_dict[cont_id] = content
    return content_dict,lexcion

# 将MFCC和number整理为能够被模型所训练的格式
def get_batch(x, y, train=False, max_pred_len=50, input_length=500):
    X = np.expand_dims(x, axis=3)
    X = x # for model2
#     labels = np.ones((y.shape[0], max_pred_len)) *  -1 # 3 # , dtype=np.uint8
    labels = y
    input_length = np.ones([x.shape[0], 1]) * ( input_length - 2 )
#     label_length = np.ones([y.shape[0], 1])
    label_length = np.sum(labels > 0, axis=1)
    label_length = np.expand_dims(label_length,1)
    inputs = {‘the_input‘: X,
              ‘the_labels‘: labels,
              ‘input_length‘: input_length,
              ‘label_length‘: label_length,
              }
    outputs = {‘ctc‘: np.zeros([x.shape[0]])}  # dummy data for dummy loss function
    return (inputs, outputs)

# 训练数据的生成器
def data_generate(wavpath = ‘D:\workspace\github\data‘, textfile = ‘D:\workspace\github\data\test.txt‘, bath_size=1):
    # 生成音频列表
    wavdict,fileids = genwavlist(wavpath)
    # 将text转化为number
    content_dict,lexcion = text2num(textfile)
    # 随意写的循环,测试编写的
    genloop = len(fileids)//bath_size
    for i in range(genloop):
        print("the ",i,"‘s loop")
        feats = []
        labels = []
        for x in range(bath_size):
            num = i * bath_size + x
            fileid = fileids[num]
            # 生成特征
            mfcc_feat = compute_mfcc(wavdict[fileid])
            # 一个batch的特征被压进来了
            feats.append(mfcc_feat)
            # 一个batch的label被压进来了
            labels.append(content_dict[fileid])
        feats = np.array(feats)
        labels = np.array(labels)
        # 整理数据格式
        inputs, outputs = get_batch(feats, labels)
        # 利用yield代替return,是生成器的特殊用法
        yield inputs, outputs

# 利用backend调用ctc
def ctc_lambda(args):
    labels, y_pred, input_length, label_length = args
    y_pred = y_pred[:, :, :]
    return K.ctc_batch_cost(labels, y_pred, input_length, label_length)


def creatModel():
    input_data = Input(name=‘the_input‘, shape=(500, 26))
    layer_h1 = Dense(512, activation="relu", use_bias=True, kernel_initializer=‘he_normal‘)(input_data)
    layer_h1 = Dropout(0.3)(layer_h1)
    layer_h2 = Dense(512, activation="relu", use_bias=True, kernel_initializer=‘he_normal‘)(layer_h1)
    layer_h3_1 = GRU(512, return_sequences=True, kernel_initializer=‘he_normal‘, dropout=0.3)(layer_h2) # GRU
    layer_h3_2 = GRU(512, return_sequences=True, go_backwards=True, kernel_initializer=‘he_normal‘, dropout=0.3)(layer_h2) # GRU
    layer_h3 = add([layer_h3_1, layer_h3_2])
    layer_h4 = Dense(512, activation="relu", use_bias=True, kernel_initializer=‘he_normal‘)(layer_h3)
    layer_h4 = Dropout(0.3)(layer_h4)
    layer_h5 = Dense(1200, activation="relu", use_bias=True, kernel_initializer=‘he_normal‘)(layer_h4)
    output = Activation(‘softmax‘, name=‘Activation0‘)(layer_h5)
    model_data = Model(inputs=input_data, outputs=output)
    #ctc
    labels = Input(name=‘the_labels‘, shape=[50], dtype=‘float32‘)
    input_length = Input(name=‘input_length‘, shape=[1], dtype=‘int64‘)
    label_length = Input(name=‘label_length‘, shape=[1], dtype=‘int64‘)
    loss_out = Lambda(ctc_lambda, output_shape=(1,), name=‘ctc‘)([labels, output, input_length, label_length])
    model = Model(inputs=[input_data, labels, input_length, label_length], outputs=loss_out)
    model.summary()
    ada_d = Adadelta(lr=0.01, rho=0.95, epsilon=1e-06)
    model.compile(loss={‘ctc‘: lambda y_true, output: output}, optimizer=ada_d)
    #test_func = K.function([input_data], [output])
    print("model compiled successful!")
    return model, model_data


model, model_data = creatModel()
# 定义数据生成器
yielddatas = data_generate()
# 使用fit_generator进行训练
model.fit_generator(yielddatas,1)
model.save_weights(‘model.mdl‘)
model_data.save_weights(‘model_data.mdl‘)
#text2num(‘E:\Data\thchs30\cv.syllable.txt‘)

后续

等训练好了试试识别怎么样,目前还没训练。

用于记录学习,希望大家提出宝贵意见

以上是关于基于CTC的语音识别系统训练的主要内容,如果未能解决你的问题,请参考以下文章

基于RNN和CTC的语音识别模型,探索语境偏移解决之道

语音识别学习资源

如何解释语音识别的技术原理?

基于LSTM+CTC的不定长验证码识别

ModelArts基于densent+ctc实现不定长文本识别训练与部署

语音识别中区分性训最大似然估计的区别是啥?