wavs_to_model.py
import wave import numpy as np import osimport keras from keras.models import Sequential from keras.layers import Densenum_class = 0 # 加载的语音文件有几种类别 labsIndName=[] ## 训练集标签的名字 ["seven","stop"]# 加载数据集 和 标签[并返回标签集的处理结果] def create_datasets():global num_class # 声明全局变量wavs=[] # 训练wav文件集labels=[] # labels 和 testlabels 这里面存的值都是对应标签的下标，下标对应的名字在labsInd中testwavs=[] # 测试wav文件集testlabels=[] # 测试集标签path="./speech_commands"dirs = os.listdir(path) # 获取的是目录列表for i in dirs:print("开始加载:",i)labsIndName.append(i) # 当前分类进入到标签的名字集wavs_path=path+"\\"+itestNum=0 # 当前分类进入了测试集的有几个 ，这里暂定每个分类进100个到测试集files = os.listdir(wavs_path) # 某个目录下文件的列表for j in files:try:waveData = get_wav_mfcc(wavs_path+"\\"+j)if testNum < 100 :testwavs.append(waveData)testlabels.append(labsIndName.index(i))testNum+=1else:wavs.append(waveData)labels.append(labsIndName.index(i))except:passnum_class = len(labsIndName) # 更新全局变量wavs=np.array(wavs)labels=np.array(labels)testwavs=np.array(testwavs)testlabels=np.array(testlabels)return (wavs,labels),(testwavs,testlabels)def get_wav_mfcc(wav_path):f = wave.open(wav_path,'rb')params = f.getparams()# print("params:",params)nchannels, sampwidth, framerate, nframes = params[:4]strData = f.readframes(nframes)#读取音频，字符串格式waveData = np.fromstring(strData,dtype=np.int16)#将字符串转化为intwaveData = waveData*1.0/(max(abs(waveData)))#wave幅值归一化waveData = np.reshape(waveData,[nframes,nchannels]).Tf.close()### 对音频数据进行长度大小的切割，保证每一个的长度都是一样的【因为训练文件全部是1秒钟长度，16000帧的，所以这里需要把每个语音文件的长度处理成一样的】data = list(np.array(waveData[0]))# print(len(data))while len(data)>16000:del data[len(waveData[0])-1]del data[0]# print(len(data))while len(data)<16000:data.append(0)# print(len(data))data=np.array(data)# 平方之后，开平方，取正数，值的范围在 0-1 之间data = data ** 2data = data ** 0.5return dataif __name__ == '__main__':(wavs,labels),(testwavs,testlabels) = create_datasets()print(wavs.shape," ",labels.shape)print(testwavs.shape," ",testlabels.shape)# 标签转换为独热码labels = keras.utils.to_categorical(labels, num_class)testlabels = keras.utils.to_categorical(testlabels, num_class)print(labels[0]) ## 类似 [1. 0]print(testlabels[0]) ## 类似 [0. 0]print(wavs.shape," ",labels.shape)print(testwavs.shape," ",testlabels.shape)# 构建模型model = Sequential()model.add(Dense(1024, activation='relu',input_shape=(16000,)))model.add(Dense(512, activation='relu'))model.add(Dense(256, activation='relu'))model.add(Dense(128, activation='relu'))model.add(Dense(num_class, activation='softmax'))# [编译模型] 配置模型，损失函数采用交叉熵，优化采用Adadelta，将识别准确率作为模型评估model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adadelta(), metrics=['accuracy'])# validation_data为验证集model.fit(wavs, labels, batch_size=124, epochs=10, verbose=1, validation_data=(testwavs, testlabels))# 开始评估模型效果 # verbose=0为不输出日志信息score = model.evaluate(testwavs, testlabels, verbose=0)print('Test loss:', score[0])print('Test accuracy:', score[1]) # 准确度model.save('asr_all_model_weights.h5') # 保存训练模型
