def test_batch(datapath = '/data/dataset/',
batch_size = 4):
# 准备测试数据，以及生成字典
p = get_data(datapath = datapath, read_type = 'test', batch_size = batch_size)
num2word = p.label_dict
yielddatas = p.data_generator()
# 载入训练好的模型，并进行识别
model, model_data = creatModel()
#model.load_weights('model_cnn_full.mdl')
# 通过修改steps增减测试数据
result = model_data.predict_generator(yielddatas, steps=2)
#print(result.shape)
pres = []
for subresult in result:
subresult = subresult.reshape((1,subresult.shape[0],subresult.shape[1]))
pre, text = decode_ctc(subresult, num2word)
#print(text)
pres.append(pre)
#print(pres)
# 获得识别结果，通过将每个识别结果与label进行比对，获得总的识别率
q = get_data(datapath = datapath, read_type = 'test', batch_size = 1)
label_gen = q.label_generator()
total_len = 0
total_err = 0
for pre in pres:
label = label_gen.__next__()
total_len += len(label)
total_err += GetEditDistance(pre, label)
print('word error rate is :', total_err/total_len*100, '%')
def train(datapath='data/', batch_size=4, steps_per_epoch=1000, epochs=1):
# 准备训练所需数据
p = get_data(datapath=datapath, read_type='train', batch_size=batch_size)
yielddatas = p.data_generator()
# 导入模型结构，训练模型，保存模型参数
model, model_data = creatModel()
if os.path.exists('speech_model/model_cnn_fbank.mdl'):
model.load_weights('speech_model/model_cnn_fbank.mdl')
model.fit_generator(yielddatas, steps_per_epoch=steps_per_epoch, epochs=1)
model.save_weights('speech_model/model_cnn_fbank.mdl')
def train(datapath = 'E:/my_ch_speech_recognition/acoustic_model/data/',
batch_size = 2,
steps_per_epoch = 1000,
epochs = 1):
# 准备训练所需数据
p = get_data(datapath = datapath, read_type = 'train', batch_size = batch_size)
yielddatas = p.data_generator()
# 导入模型结构，训练模型，保存模型参数
model, model_data = creatModel()
if os.path.exists('model_cnn_full.mdl'):
model.load_weights('model_cnn_full.mdl')
model.fit_generator(yielddatas, steps_per_epoch=steps_per_epoch, epochs=1)
model.save_weights('model_cnn_full.mdl')
def test(datapath='data/', batch_size=1):
# 准备测试数据，以及生成字典
p = get_data(datapath=datapath, read_type='test', batch_size=batch_size)
num2word = p.label_dict
yielddatas = p.data_generator()
# 载入训练好的模型，并进行识别
model, model_data = creatModel()
model.load_weights('speech_model/model_cnn_fbank.mdl')
result = model_data.predict_generator(yielddatas, steps=1)
print(result.shape)
# 将数字结果转化为文本结果
result, text = decode_ctc(result, num2word)
print('数字结果： ', result)
print('文本结果：', text)
