From fb89aabbad9c9786c889998d3ed0f738c711070f Mon Sep 17 00:00:00 2001
From: nl8590687 <3210346136@qq.com>
Date: Fri, 11 May 2018 16:56:59 +0800
Subject: [PATCH] fix bugs and improve asrserver
---
.gitignore | 1 +
SpeechModel22.py | 49 +++++++++++++++++++++++++++++-----
asrserver.py | 20 ++++++--------
dict.txt | 24 ++++++++---------
general_function/file_wav.py | 51 +++++++++++++++++++++++++++++++-----
test.py | 2 +-
6 files changed, 109 insertions(+), 38 deletions(-)
diff --git a/.gitignore b/.gitignore
index a61d32b..cb5eb65 100644
--- a/.gitignore
+++ b/.gitignore
@@ -5,6 +5,7 @@
__pycache__
*.wav
+*.model_yaml
Test_Report_*
dataset
diff --git a/SpeechModel22.py b/SpeechModel22.py
index 01f79b3..a3f905c 100644
--- a/SpeechModel22.py
+++ b/SpeechModel22.py
@@ -16,6 +16,7 @@
import numpy as np
import random
+from keras.models import model_from_yaml
from keras.models import Sequential, Model
from keras.layers import Dense, Dropout, Input, Reshape # , Flatten,LSTM,Convolution1D,MaxPooling1D,Merge
from keras.layers import Conv1D,LSTM,MaxPooling1D, Lambda, TimeDistributed, Activation,Conv2D, MaxPooling2D #, Merge,Conv1D
@@ -40,7 +41,14 @@ def __init__(self, datapath):
self.label_max_string_length = 64
self.AUDIO_LENGTH = 1600
self.AUDIO_FEATURE_LENGTH = 200
+
+ self.model_name = 'm22'
+
+ #if(not os.path.exists(self.model_name + '.model_yaml')): # 判断保存模型的目录是否存在
self._model, self.base_model = self.CreateModel()
+ #else:
+ # self._model, self.base_model = self.load_model_yaml(self.model_name)
+
self.datapath = datapath
self.slash = ''
@@ -112,7 +120,7 @@ def CreateModel(self):
model = Model(inputs=[input_data, labels, input_length, label_length], outputs=loss_out)
- model.summary()
+ #model.summary()
# clipnorm seems to speeds up convergence
#sgd = SGD(lr=0.0001, decay=1e-6, momentum=0.9, nesterov=True, clipnorm=5)
@@ -125,6 +133,9 @@ def CreateModel(self):
# captures output of softmax so we can decode the output during visualization
test_func = K.function([input_data], [y_pred])
+ #kr.utils.plot_model(model, to_file='model.png', show_shapes=False, show_layer_names=True) # 可视化展示模型
+ self.save_model_yaml(model, model_data)
+
print('[*提示] 创建模型成功，模型编译成功')
return model, model_data
@@ -135,7 +146,32 @@ def ctc_lambda_func(self, args):
#y_pred = y_pred[:, 2:, :]
return K.ctc_batch_cost(labels, y_pred, input_length, label_length)
+ def save_model_yaml(self,model,model_data):
+ '''
+ 保存模型的配置结构
+ '''
+ str_yaml_model = model.to_yaml()
+ str_yaml_model_data = model_data.to_yaml()
+ f = open(self.model_name + '.model_yaml','w')
+ f.write(str_yaml_model)
+ f.close()
+ f = open(self.model_name + '_base.model_yaml','w')
+ f.write(str_yaml_model_data)
+ f.close()
+ def load_model_yaml(self, model_name):
+ '''
+ 加载模型的配置结构
+ '''
+ f = open(self.model_name + '.model_yaml','r')
+ str_yaml_model = f.read()
+ f.close()
+ f = open(self.model_name + '_base.model_yaml','r')
+ str_yaml_model_data = f.read()
+ f.close()
+ model = model_from_yaml(str_yaml_model)
+ model_data = model_from_yaml(str_yaml_model_data)
+ return model, model_data
def TrainModel(self, datapath, epoch = 2, save_step = 1000, batch_size = 32, filename = 'model_speech/speech_model2'):
'''
@@ -230,7 +266,7 @@ def TestModel(self, datapath='', str_dataset='dev', data_count = 32, out_report
txt += 'True:\t' + str(data_labels) + '\n'
txt += 'Pred:\t' + str(pre) + '\n'
txt += '\n'
- txt_obj.write(txt)
+ txt_obj.write(txt)
print('*[测试结果] 语音识别 ' + str_dataset + ' 集语音单字错误率：', word_error_num / words_num * 100, '%')
if(out_report == True):
@@ -356,7 +392,7 @@ def model(self):
import tensorflow as tf
from keras.backend.tensorflow_backend import set_session
- os.environ["CUDA_VISIBLE_DEVICES"] = "1"
+ os.environ["CUDA_VISIBLE_DEVICES"] = "0"
#进行配置，使用70%的GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.7
@@ -385,9 +421,10 @@ def model(self):
ms = ModelSpeech(datapath)
- #ms.LoadModel(modelpath + 'm22_2\\1\\speech_model22_e_0_step_159000.model')
- ms.TrainModel(datapath, epoch = 50, batch_size = 4, save_step = 500)
- #ms.TestModel(datapath, str_dataset='test', data_count = 128, out_report = True)
+ #ms.LoadModel(modelpath + 'm22_2\\1\\speech_model22_e_0_step_327500.model')
+ ms.LoadModel(modelpath + 'm22_2/1/speech_model22_e_0_step_327500.model')
+ #ms.TrainModel(datapath, epoch = 50, batch_size = 4, save_step = 500)
+ #ms.TestModel(datapath, str_dataset='train', data_count = 128, out_report = True)
#r = ms.RecognizeSpeech_FromFile('E:\\语音数据集\\ST-CMDS-20170001_1-OS\\20170001P00241I0053.wav')
#r = ms.RecognizeSpeech_FromFile('E:\\语音数据集\\ST-CMDS-20170001_1-OS\\20170001P00020I0087.wav')
#r = ms.RecognizeSpeech_FromFile('E:\\语音数据集\\wav\\train\\A11\\A11_167.WAV')
diff --git a/asrserver.py b/asrserver.py
index d047ff4..bfa7250 100644
--- a/asrserver.py
+++ b/asrserver.py
@@ -11,6 +11,13 @@
from SpeechModel22 import ModelSpeech
from LanguageModel import ModelLanguage
+datapath = 'data/'
+modelpath = 'model_speech/'
+ms = ModelSpeech(datapath)
+ms.LoadModel(modelpath + 'speech_model22_e_0_step_216500.model')
+
+ml = ModelLanguage('model_language')
+ml.LoadModel()
class TestHTTPHandle(http.server.BaseHTTPRequestHandler):
@@ -66,7 +73,7 @@ def do_POST(self):
if(token == 'qwertasd'):
#buf = '成功\n'+'wavs:\n'+str(wavs)+'\nfs:\n'+str(fs)
- buf = r[0]
+ buf = r
else:
buf = '403'
@@ -74,24 +81,13 @@ def do_POST(self):
self._set_response()
-
-
-
-
#buf = ' \n \n\nPost page\n \nPost Data:%s
Path:%s\n \n'%(datas,self.path)
buf = bytes(buf,encoding="utf-8")
self.wfile.write(buf)
def recognize(self, wavs, fs):
- datapath = 'data/'
- modelpath = 'model_speech/'
- ms = ModelSpeech(datapath)
- ms.LoadModel(modelpath + 'speech_model22_e_0_step_6500.model')
-
r_speech = ms.RecognizeSpeech(wavs, fs)
- ml = ModelLanguage('model_language')
- ml.LoadModel()
str_pinyin = r_speech
r = ml.SpeechToText(str_pinyin)
return r
diff --git a/dict.txt b/dict.txt
index 160efb0..0a692c7 100644
--- a/dict.txt
+++ b/dict.txt
@@ -1,13 +1,13 @@
-a1 阿啊呵腌吖锕雅
+a1 阿啊呵腌吖锕
a2 啊呵嗄
a3 啊呵
a4 啊呵
-a5 阿啊呵娃
+a5 阿啊呵
ai1 哀挨埃唉哎捱锿诶
ai2 呆挨癌皑捱矮
ai3 矮哎蔼霭嗳
-ai4 爱碍艾唉哎隘暧嗳瑷嗌嫒砹愛以
-an1 安谙鞍氨庵桉鹌广厂
+ai4 爱碍艾唉哎隘暧嗳瑷嗌嫒砹愛
+an1 安谙鞍氨庵桉鹌
an3 俺铵揞埯
an4 案按暗岸黯胺犴
ang1 肮
@@ -19,7 +19,7 @@ ao3 袄拗媪
ao4 奥澳傲懊坳拗骜岙鏊
ba1 八吧巴叭芭扒疤笆粑岜捌
ba2 八拔跋茇菝魃
-ba3 把靶钯靶星
+ba3 把靶钯靶
ba4 把爸罢霸坝耙灞鲅壩
ba5 吧罢巴叭
bai1 掰
@@ -34,11 +34,11 @@ bang3 膀榜绑
bang4 棒膀傍磅谤镑蚌蒡
bao1 包胞炮剥褒苞孢煲龅
bao2 薄雹保
-bao3 保宝饱堡葆褓鸨乖
-bao4 报暴抱爆鲍曝刨瀑豹趵在
+bao3 保宝饱堡葆褓鸨
+bao4 报暴抱爆鲍曝刨瀑豹趵
bei1 背悲杯碑卑陂埤萆鹎
bei3 北
-bei4 被备背辈倍贝蓓惫悖狈焙邶钡孛碚褙鐾鞴宝
+bei4 被备背辈倍贝蓓惫悖狈焙邶钡孛碚褙鐾鞴
bei5 臂呗备
ben1 奔贲锛
ben3 本苯畚
@@ -82,7 +82,7 @@ cai1 猜
cai2 才财材裁采
cai3 采彩踩睬
cai4 采菜蔡
-can1 参餐骖食
+can1 参餐骖
can2 残惭蚕
can3 惨黪
can4 惨灿掺璨孱粲
@@ -95,7 +95,7 @@ ce4 策测侧厕册恻
cen1 参
cen2 岑涔
ceng1 噌
-ceng2 曾层太
+ceng2 曾层
ceng4 蹭
cha1 差插叉碴喳嚓杈馇锸
cha2 查察茶叉茬碴楂猹搽槎檫
@@ -484,8 +484,8 @@ kao3 考烤拷栲
kao4 靠铐犒
ke1 科颗柯呵棵苛磕坷嗑瞌轲稞疴蝌钶窠颏珂髁
ke2 咳壳颏可
-ke3 可渴坷轲岢以
-ke4 可克客刻课恪嗑溘骒缂氪锞蚵科谎
+ke3 可渴坷轲岢
+ke4 克客刻课恪嗑溘骒缂氪锞蚵科可
ken3 肯恳啃垦龈
ken4 裉
keng1 坑吭铿
diff --git a/general_function/file_wav.py b/general_function/file_wav.py
index d64b36d..b1c292d 100644
--- a/general_function/file_wav.py
+++ b/general_function/file_wav.py
@@ -102,6 +102,37 @@ def GetFrequencyFeature2(wavsignal, fs):
#print(data_input.shape)
return data_input
+def GetFrequencyFeature3(wavsignal, fs):
+ # wav波形 加时间窗以及时移10ms
+ time_window = 25 # 单位ms
+ window_length = fs / 1000 * time_window # 计算窗长度的公式，目前全部为400固定值
+
+ wav_arr = np.array(wavsignal)
+ #wav_length = len(wavsignal[0])
+ wav_length = wav_arr.shape[1]
+
+ range0_end = int(len(wavsignal[0])/fs*1000 - time_window) // 10 # 计算循环终止的位置，也就是最终生成的窗数
+ data_input = np.zeros((range0_end, 200), dtype = np.float) # 用于存放最终的频率特征数据
+ data_line = np.zeros((1, 400), dtype = np.float)
+ for i in range(0, range0_end):
+ p_start = i * 160
+ p_end = p_start + 400
+
+ data_line = wav_arr[0, p_start:p_end]
+
+ x=np.linspace(0, 400 - 1, 400, dtype = np.int64)
+ w = 0.54 - 0.46 * np.cos(2 * np.pi * (x) / (400 - 1) ) # 汉明窗
+ data_line = data_line * w # 加窗
+
+ data_line = np.abs(fft(data_line)) / wav_length
+
+
+ data_input[i]=data_line[0:200] # 设置为400除以2的值（即200）是取一半数据，因为是对称的
+
+ #print(data_input.shape)
+ data_input = np.log(data_input + 1)
+ return data_input
+
def wav_scale(energy):
'''
语音信号能量归一化
@@ -178,13 +209,19 @@ def get_wav_symbol(filename):
return dic_symbol_list,list_symbolmark
if(__name__=='__main__'):
- #dic=get_wav_symbol('E:\\语音数据集\\doc\\doc\\trans\\train.syllable.txt')
- #print(dic)
- #dic=get_wav_list('E:\\语音数据集\\doc\\doc\\list\\train.wav.lst')
- #for i in dic:
- #print(i,dic[i])
+
wave_data, fs = read_wav_data("A2_0.wav")
- #wave_data[0]=wav_scale(wave_data[0])
- #print(fs)
+
wav_show(wave_data[0],fs)
+ #t0=time.time()
+ freimg = GetFrequencyFeature3(wave_data,fs)
+ #t1=time.time()
+ #print('time cost:',t1-t0)
+
+ freimg = freimg.T
+ plt.subplot(111)
+ plt.imshow(freimg)
+ plt.colorbar(cax=None,ax=None,shrink=0.5)
+
+ plt.show()
diff --git a/test.py b/test.py
index 1618847..a7edd95 100644
--- a/test.py
+++ b/test.py
@@ -32,7 +32,7 @@
ms.LoadModel(modelpath + 'm22_2/0/speech_model22_e_0_step_257000.model')
#ms.TestModel(datapath, str_dataset='test', data_count = 64, out_report = True)
-r = ms.RecognizeSpeech_FromFile('E:\语音数据集\ST-CMDS-20170001_1-OS\\20170001P00241I0052.wav')
+r = ms.RecognizeSpeech_FromFile('E:\\语音数据集\\ST-CMDS-20170001_1-OS\\20170001P00241I0052.wav')
#r = ms.RecognizeSpeech_FromFile('E:\语音数据集\ST-CMDS-20170001_1-OS\\20170001P00241I0053.wav')
#r = ms.RecognizeSpeech_FromFile('E:\\语音数据集\\ST-CMDS-20170001_1-OS\\20170001P00020I0087.wav')
#r = ms.RecognizeSpeech_FromFile('E:\\语音数据集\\wav\\train\\A11\\A11_167.WAV')