from pyAudioAnalysis import audioBasicIO

from pyAudioAnalysis import audioFeatureExtraction

import numpy as np

import math

import matplotlib.pyplot as plt

# main process

[Fs, x] = audioBasicIO.readAudioFile("data/diarizationExample.wav")

TIME_OF_WINDOW = 0.050 #a window = 0.05s

TIME_OF_STEP = 0.025 #step = 0.01s

SIZE_OF_WINDOW = int(TIME_OF_WINDOW * Fs) #the number of frame for one window

SIZE_OF_STEP = int(TIME_OF_STEP * Fs) #the number of frame for one step

BLOCK_SIZE = 4 #a block has (6 * SIZE_OF_STEP) frame

BLOCK_STEP = 2

# variables

END_OF_FILE = 0

FIRST_PAIR = 1

INDEX_BOUCLE = 1

def getMFCCs(block_start, block_end):

return attribute[8:20,block_start:block_end+1]

def getMFCCsFromTime(moment_start, moment_end):

block_start = int(moment_start / BLOCK_STEP / TIME_OF_STEP - 1)

block_end = int(moment_end / BLOCK_STEP / TIME_OF_STEP - 1)

return getMFCCs(block_start, block_end)

def gauss(x, mean, cov):

return y

# feature extraction from the library pyAudioAnalysis

attribute = audioFeatureExtraction.stFeatureExtraction(x, Fs, SIZE_OF_WINDOW, SIZE_OF_STEP)

# relationship between the similarity and the timestamp in the audio

relation = [[1 for col in range(2)] for row in range(attribute.shape[1]/BLOCK_STEP)]

while (END_OF_FILE == 0):

if (FIRST_PAIR == 1):

# for the first pari of the block

block_i_index_start = 0

block_i_index_end = BLOCK_SIZE

block_i_attribute = getMFCCs(block_i_index_start, block_i_index_end)

block_i_mean = np.mean(block_i_attribute, axis=1)

block_i_cov = np.cov(block_i_attribute)

block_i_log_like = np.log(gauss(block_i_attribute, mean=block_i_mean, cov=block_i_cov))

block_j_index_start = block_i_index_end + 1

block_j_index_end = block_j_index_start + BLOCK_SIZE - 1

block_j_attribute = getMFCCs(block_j_index_start, block_j_index_end)

block_j_mean = np.mean(block_j_attribute, axis=1)

block_j_cov = np.cov(block_j_attribute)

block_j_log_like = np.log(gauss(block_j_attribute, mean=block_j_mean, cov=block_j_cov))

FIRST_PAIR = 0

else:

#for the rest of the block

block_j_index_start += BLOCK_STEP

block_j_index_end += BLOCK_STEP

new_attribute = getMFCCs(block_j_index_end-BLOCK_STEP+1, block_j_index_end)

#the following code is for the object that to avoid recalculate the overlap between the block after moved and before moved

block_i_index_start += BLOCK_STEP

block_i_index_end += BLOCK_STEP

block_i_attribute[:,0:block_i_attribute.shape[1]-new_attribute.shape[1]] = block_i_attribute[:,new_attribute.shape[1]:block_i_attribute.shape[1]]

block_i_attribute[:,block_i_attribute.shape[1]-new_attribute.shape[1]:block_i_attribute.shape[1]] = block_j_attribute[:,0:new_attribute.shape[1]]

block_i_mean = np.mean(block_i_attribute, axis=1)

block_i_cov = np.cov(block_i_attribute)

block_i_log_like = np.log(gauss(block_i_attribute, mean=block_i_mean, cov=block_i_cov))

block_j_attribute[:,0:block_j_attribute.shape[1]-new_attribute.shape[1]] = block_j_attribute[:,new_attribute.shape[1]:block_j_attribute.shape[1]]

block_j_attribute[:,block_j_attribute.shape[1]-new_attribute.shape[1]:block_j_attribute.shape[1]] = new_attribute[:,0:new_attribute.shape[1]]

block_j_mean = np.mean(block_j_attribute, axis=1)

block_j_cov = np.cov(block_j_attribute)

block_j_log_like = np.log(gauss(block_j_attribute, mean=block_j_mean, cov=block_j_cov))

block_union_index_start = block_i_index_start

block_union_index_end = block_j_index_end

block_union_attribute = np.concatenate((block_i_attribute, block_j_attribute), axis = 1)

block_union_mean = np.mean(block_union_attribute, axis=1)

block_union_cov = np.cov(block_union_attribute)

block_union_log_like = np.log(gauss(block_union_attribute, mean=block_union_mean, cov=block_union_cov))

relation[INDEX_BOUCLE-1][0] = np.sum(block_i_log_like) + np.sum(block_j_log_like) - np.sum(block_union_log_like)

relation[INDEX_BOUCLE-1][1] = (block_i_index_end + block_j_index_start) / 2 * TIME_OF_STEP

INDEX_BOUCLE += 1

if block_j_index_end + BLOCK_STEP > attribute.shape[1]:

END_OF_FILE = 1

# cut the audio

relation_cut = filter(lambda t: t[0] > 0, relation)

print "=========The data after cutting==========="

for item in relation_cut:

print item

# print getMFCCsFromTime(0.4, 4.0).shape

#===========Kmeans Definition=====================

#calculate Euclidean distance

def euclDistance(vector1, vector2):

return sqrt(sum(power(vector2 - vector1, 2)))

# init centroids with random samples

def initCentroids(dataSet, k):

return centroids

# k-means cluster

def kmeans(dataSet, k):

return centroids, clusterAssment

def abs (input):

if input <0 : return - input

else : return input

print '======='

temprelation = []

for i in range(len(relation_cut)-3): #-3 because the last 3 numbers are meaningless

temprelation.append([relation_cut[i][0],1])

#temprelation is a list which takes only first colum of relation_cut and add another colum which values are all 1

#This action alow to transefer the MFCC to a 2-dimension data which use for cluster

## step 1: load data

print "step 1: load data..."

dataSet = temprelation

# step 2: clustering...

print "step 2: clustering..."

dataSet = mat(dataSet)

k = 4

centroids, clusterAssment = kmeans(dataSet, k)

classification =[]

for indexi in range(k):

classification .append(centroids[indexi][0])

## step 3: Mark speaker to each block

relation_cut_regonize = [] #This list is to store the information with different speakers

for indexj in range(len(dataSet)):

min1 = abs(relation_cut[indexj][0]-classification[0])

min2 = abs(relation_cut[indexj][0]-classification[1])

min3 = abs(relation_cut[indexj][0]-classification[2])

min4 = abs(relation_cut[indexj][0]-classification[3])

if min1<min2 and min1 < min3 and min1<min4:

relation_cut_regonize.append([relation_cut[indexj],"Speaker1"])

else:

if min2<min1 and min2 < min3 and min2<min4:

relation_cut_regonize.append([relation_cut[indexj],"Speaker2"])

else:

if min3<min1 and min3 < min2 and min3<min4:

relation_cut_regonize.append([relation_cut[indexj],"Speak3"])

else:

if min4<min1 and min4 < min2 and min4<min3:

relation_cut_regonize.append([relation_cut[indexj],"Speak4"])

#step4: shows the result

print "=========The data after regonizing as 4 speakers==========="

for item in relation_cut_regonize:

print item