暑假实习，作语音识别的项目，由于要用到TCP进行数据的传输，因而向同事学习了如何实现服务端和客户端数据传输。记下来以便之后能够用。python
话很少说，上菜：数组
（注：只是用来实现语音数据的一个简单传输，勿喷！）socket
client：函数
import socket
import argparse
HOST = '192.168.1.100'
PORT = 88
def parse_args():
'''使用argparse的第一步是建立一个 ArgumentParser对象,
建立一个 ArgumentParser对象，这个ArgumentParser对象中会保存
全部将命令行参数转为python数据类型的必需信息
'''
#使用 argparse.ArgumentParser建立ArgumentParser对象
parser = argparse.ArgumentParser()
#每个add_argument方法对应一个你要关注的参数或选项
parser.add_argument('--file', type=str, required=True,
help='path to wav file')
return parser.parse_args()
def main(args):
#建立socker实例
s = socket.socket()
##客户端创建链接,host是回环地址，通常用于测试
s.connect((HOST, PORT))
print('connected with server')
f = open(args.file, 'rb')
print('sending file')
data = f.read(1024)
while data:
s.send(data)
#读取1024字节
data = f.read(1024)
'''
使用shutdown来关闭socket的功能
SHUT_RDWR：关闭读写，即不能使用send/write/recv/read等
SHUT_RD：关闭读，即不能使用read/recv等
SHUT_WR:关闭写功能，即不能使用send/write等
'''
s.shutdown(socket.SHUT_WR)
f.close()
print('done sending data...waiting for response')
print('识别结果：' + s.recv(1024).decode('utf-8'))
if __name__ == '__main__':
args = parse_args()
main(args)
sever：学习
import socket
import numpy as np
import tensorflow as tf
#读取.wav音频文件.它会返回一个元组,第一项为音频的采样率,第二项为音频数据的numpy数组.
import scipy.io.wavfile as wav
#该模块的主要功能有：语音识别、将指定文本合成语音以及语音信号输出等
import speechpy
import utils
from model_helper import las_model_fn
VOCAB_DIR = 'output/tfdata/vocab.table'
MODEL_DIR = 'output/model'
SAVE_PATH = 'bin/'
BEAM_WIDTH = 3
PORT = 8888
def input_fn(dataset_filename, vocab_filename, num_channels=39, batch_size=8, num_epochs=1):
dataset = utils.read_dataset(dataset_filename, num_channels)
vocab_table = utils.create_vocab_table(vocab_filename)
dataset = utils.process_dataset(
dataset, vocab_table, utils.SOS, utils.EOS, batch_size, num_epochs, is_infer=True)
return dataset
def extract_mfcc(filename):
if not filename.endswith('.wav'): #用于判断字符串是否以指定后缀结尾
return None, None
fs, signal = wav.read(filename)
print('frequency={0}'.format(fs))
#assert fs == 48000
# downsample
signal = signal[::3]
fs = 16000
#从音频信号计算MFCC功能，返回包含mfcc特征的numpy数组
mfcc = speechpy.feature.mfcc(signal, fs)
#此功能旨在执行全球倒谱平均值和输入特征向量“vec”上的方差归一化（CMVN）
#返回： 均值（或均值+方差）归一化特征向量
mfcc_cmvn = speechpy.processing.cmvn(mfcc, True)
#该函数提取时间导数特征，返回排列
mfcc_39 = speechpy.feature.extract_derivative_feature(mfcc_cmvn)
return filename[:-4], mfcc_39.reshape(-1, 39)
def predict(args, model):
predictions = model.predict(
input_fn=lambda: input_fn(
args.data, args.vocab, num_channels=args.num_channels, batch_size=args.batch_size, num_epochs=1),
predict_keys='sample_ids')
if args.beam_width > 0:
predictions = [vocab_list[y['sample_ids'][:, 0]].tolist() + [utils.EOS]
for y in predictions]
else:
predictions = [vocab_list[y['sample_ids']].tolist() + [utils.EOS]
for y in predictions]
predictions = [y[:y.index(utils.EOS)] for y in predictions]
predictions = predictions[0]
# with open(args.save, 'w+') as f:
# f.write('\n'.join(predictions))
# predictions = ''.join(predictions)
result = ''
for c in predictions:
print('c=',c)
result += c
return result
class Args:
def __init__(self):
self.vocab = VOCAB_DIR
self.model_dir = MODEL_DIR
self.beam_width = BEAM_WIDTH
self.save = SAVE_PATH
self.batch_size = 8
self.num_channels = 39
if __name__ == '__main__':
# tf.logging.set_verbosity(tf.logging.INFO)
print("initializing LAS model...")
args = Args()
vocab_list = np.array(utils.load_vocab(args.vocab))
vocab_size = len(vocab_list)
config = tf.estimator.RunConfig(model_dir=args.model_dir)
hparams = utils.create_hparams(
args, vocab_size, utils.SOS_ID, utils.EOS_ID)
hparams.decoder.set_hparam('beam_width', args.beam_width)
model = tf.estimator.Estimator(
model_fn=las_model_fn,
config=config,
params=hparams)
# socket connection
skt = socket.socket()
#绑定端口
skt.bind(('', PORT))
''' listen等待客户端链接--开始TCP监听。backlog指定在拒绝链接以前，操做系统能够挂起的最大连
接数量。该值至少为1，大部分应用程序设为5就能够了。
'''
skt.listen()
print('server listening on port ', PORT)
while True:
#调用socket对象的accept方法.该方法等待客户端的链接,并返回connection对象,表示已链接到客户端
conn, addr = skt.accept()
with conn:
print('client {} connected...'.format(addr))
with open('bin/received.wav', 'wb') as f:
while True:
print('receiving data...')
data = conn.recv(1024)
if not data:
print('no data...')
f.close()
break
# write data to a file
print('writing data...')
f.write(data)
# finished file transfering, now process file
print('processing data...')
try:
name, feature = extract_mfcc('bin/received.wav')
except:
conn.send('ERROR:语音文件格式不匹配'.encode('utf-8'))
continue
feature_lists = tf.train.FeatureLists(feature_list={
'labels': tf.train.FeatureList(feature=[
tf.train.Feature(bytes_list=tf.train.BytesList(value=[''.encode()]))
]),
'inputs': tf.train.FeatureList(feature=[
tf.train.Feature(float_list=tf.train.FloatList(value=f))
for f in feature
])
})
with tf.python_io.TFRecordWriter('{0}.tfrecord'.format(name)) as writer:
writer.write(tf.train.SequenceExample(feature_lists=feature_lists).SerializeToString())
args.data = '{0}.tfrecord'.format(name)
print('inferring...')
s = predict(args, model)
conn.send(s.encode('utf-8'))
conn.close()
os.system
路很长，要加油哦~~~测试