In[2]:

#import libraries import librosa get_ipython().run_line_magic('pylab', 'inline') import os import pandas as pd import librosa.display import glob import IPython.display # display audio data as sound import keras

In[3]:

get_ipython().system('pip install data_utils')

In[4]:

#Load the audio data file data, sampling_rate = librosa.load('E:/gaurav_back up/Spring 2018/voice_recognition/train/Train/2022.wav')

In[5]:

What is the numerical output of a audio file?

data,sampling_rate

# STEP1: Visualization

In[6]:

#By hearing import IPython.display as ipd ipd.Audio('E:/gaurav_back up/Spring 2018/voice_recognition/train/Train/2022.wav')

In[7]:

#By graphs plt.figure(figsize=(12, 4)) librosa.display.waveplot(data, sr=sampling_rate)

STEP 2: Insights about data.

In[8]:

Setting directories becuase we have 5435 .wav,individual files to read.

root_dir = os.path.abspath('.') data_dir_train = 'E:/gaurav_back up/Spring 2018/voice_recognition/train/Train/' train = pd.read_csv('E:/gaurav_back up/Spring 2018/voice_recognition/train/train.csv') data_dir_test = 'E:/gaurav_back up/Spring 2018/voice_recognition/test/Test/' test = pd.read_csv('E:/gaurav_back up/Spring 2018/voice_recognition/test/test.csv')

In[34]:

#Magic of randamization i = random.choice(train.index) print(i) audio_name = train.ID[i] path = os.path.join(data_dir_train, 'Train', str(audio_name) + '.wav')

print('Class: ', train.Class[i]) x, sr = librosa.load(data_dir_train + str(train.ID[i]) + '.wav')

plt.figure(figsize=(12, 4)) librosa.display.waveplot(x, sr=sr)

Step 3: Let's have out first model. Majority wins

In[10]:

(train.Class.value_counts())

In[11]:

train.Class.value_counts()[:].plot(kind='barh')

In[12]:

#Model 1: Concept of majority

a daily life model. it is averaging or just labeling every one based on mazority

test['Class'] ='jackhammer' test.to_csv('sub01.csv',index=False)

STEP 4: Feature extraction

In[13]:

def parser(row):

function to load files and extract features

data_dir = 'E:/gaurav_back up/Spring 2018/voice_recognition/train/' file_name = os.path.join(os.path.abspath(data_dir),'Train',str(row.ID) + '.wav') print(file_name)

handle exception to check if there isn't a file which is corrupted

try: X, sample_rate = librosa.load(file_name, res_type='kaiser_fast') # we extract mfcc feature from data #print(X,sample_rate) mfccs = np.mean(librosa.feature.mfcc(y=X, sr=sample_rate, n_mfcc=40).T,axis=0) except Exception as e: print("Error encountered while parsing file: ", file_name) return None

feature = mfccs label = row.Class

return [feature, label]

temp = train.apply(parser, axis=1) temp.columns = ['feature', 'label'] print(temp)

STEP 5: Convert the data to pass it in Deep learning model

In[14]:

from sklearn.preprocessing import LabelEncoder import keras import numpy

#For training set X1 = np.array(temp.feature.tolist()) X =list(filter(None.ne, X1))

#For training set y1 = np.array(temp.label.tolist()) y = [x for x in y1 if x != None]

lb = LabelEncoder()

#Label encoding for training set p=lb.fit_transform(y) print(p) y=keras.utils.to_categorical(p) print(y)

In[15]:

np.array(X).shape

In[16]:

len(X1)

In[17]:

len(X)

Step 6:Run a deep learning model and study results

In[18]:

import numpy as np from keras.models import Sequential #We can add individual layers to our neural network one layer at a time. from keras.layers import Dense, Dropout, Activation, Flatten from keras.layers.convolutional import Conv2D from keras.layers import Convolution2D, MaxPooling2D from keras.optimizers import Adam,RMSprop #from keras.utils import np_utils from sklearn import metrics

num_labels = y.shape[1] print(num_labels) filter_size = 2

model = Sequential() model.add(Dense(256, input_shape=(40,))) model.add(Activation('relu'))

model.add(Dropout(0.5)) model.add(Dense(10,activation='softmax'))

In[19]:

#(b)

optimizer function that are faster than gradient descent

#1 Momentum Optimization #2 Nesterov Accelerated Gradient #3 RMSProp() #4 Adam model.compile(loss='categorical_crossentropy', metrics=['accuracy'], optimizer='Adam')

http://playground.tensorflow.org/#activation=relu&batchSize=10&dataset=spiral&regDataset=reg-plane&learningRate=0.03&regularizationRate=0&noise=0&networkShape=6,2&seed=0.86127&showTestData=false&discretize=false&percTrainData=50&x=true&y=true&xTimesY=false&xSquared=false&ySquared=false&cosX=false&sinX=false&cosY=false&sinY=false&collectStats=false&problem=classification&initZero=false&hideText=false

In[20]:

model.summary()

Step 7 : Train the model

In[21]:

model.fit(np.array(X),np.array(y),batch_size=32,epochs=5)

In[22]:

cross_validation

model.fit(np.array(X),np.array(y),batch_size=32,epochs=10,validation_split=0.3)

we see there are better results on the validation set

Step 8 : Training to Testing

In[23]:

def parser2(row):

function to load files and extract features

data_dir = 'E:/gaurav_back up/Spring 2018/voice_recognition/test/' file_name = os.path.join(os.path.abspath(data_dir),'Test',str(row.ID) + '.wav') print(file_name)

handle exception to check if there isn't a file which is corrupted

try: X, sample_rate = librosa.load(file_name, res_type='kaiser_fast') # we extract mfcc feature from data #print(X,sample_rate) mfccs = np.mean(librosa.feature.mfcc(y=X, sr=sample_rate, n_mfcc=40).T,axis=0) except Exception as e:

  print("Error encountered while parsing file: ", file_name)
  return None

feature = mfccs #label = row.Class

return [feature]

temp1 = test.apply(parser2, axis=1)

#temp1.columns = ['ID','feature']

print(temp1)

In[24]:

print(temp1) temp1.columns = ["feature"] print(temp1)

In[26]:

#Processing the test file. Removing corrupt .wav file from sklearn.preprocessing import LabelEncoder import np_utils import keras import numpy

#For testing set X1 = np.array(temp1.tolist()) test_X =list(filter(None.ne, X1))

#For testing set p1=np.array(test)

In[44]:

np.array(temp2).shape

In[45]:

test_X = np.reshape(temp2, (2131,40),1)

In[47]:

#test_X #model1 pred = model.predict_classes((np.array(test_X))) pred = lb.inverse_transform(pred) test['Class'] = pred #test.to_csv(‘sub02.csv’, index=False)

In[30]:

pred

In[39]:

temp2 = [t for t in temp1 if t is not None]

In[42]:

temp2

key take aways:

Nomalization is important as neural network usually work better than. It has mean of 0 and unit variance.

#scikit_learn's StandardScaler can do this,but do not forget to use it.